Psychopharmacology

https://doi.org/10.1007/s00213-020-05577-x

REVIEW

Rage Against the Machine: Advancing the study of aggression

ethology via machine learning.
Nastacia L. Goodwin 1,2 & Simon R.O. Nilsson 1 & Sam A. Golden 1,2,3

Received: 7 November 2019 / Accepted: 1 June 2020

# Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Rationale Aggression, comorbid with neuropsychiatric disorders, exhibits with diverse clinical presentations and places a sig-
nificant burden on patients, caregivers, and society. This diversity is observed because aggression is a complex behavior that can
be ethologically demarcated as either appetitive (rewarding) or reactive (defensive), each with its own behavioral characteristics,
functionality, and neural basis that may transition from adaptive to maladaptive depending on genetic and environmental factors.
There has been a recent surge in the development of preclinical animal models for studying appetitive aggression-related
behaviors and identifying the neural mechanisms guiding their progression and expression. However, adoption of these proce-
dures is often impeded by the arduous task of manually scoring complex social interactions. Manual observations are generally
susceptible to observer drift, long analysis times, and poor inter-rater reliability, and are further incompatible with the sampling
frequencies required of modern neuroscience methods.
Objectives In this review, we discuss recent advances in the preclinical study of appetitive aggression in mice, paired with our
perspective on the potential for machine learning techniques in producing automated, robust scoring of aggressive social
behavior. We discuss critical considerations for implementing valid computer classifications within behavioral pharmacological
studies.
Key results Open-source automated classification platforms can match or exceed the performance of human observers while
removing the confounds of observer drift, bias, and inter-rater reliability. Furthermore, unsupervised approaches can identify
previously uncharacterized aggression-related behavioral repertoires in model species.
Discussion and conclusions Advances in open-source computational approaches hold promise for overcoming current manual
annotation caveats while also introducing and generalizing computational neuroethology to the greater behavioral neuroscience
community. We propose that currently available open-source approaches are sufficient for overcoming the main limitations
preventing wide adoption of machine learning within the context of preclinical aggression behavioral research.

Keywords Aggression . Machine learning . Mice . Motivation . Reward

Introduction

There is an increased risk for abnormal or pathological aggres-

Nastacia L. Goodwin and Simon R.O. Nilsson contributed equally to this sion in individuals suffering from neuropsychiatric disorders.
work. A growing literature indicates that certain types of aggression,
namely compulsive aggression seeking, may be a distinct ex-
* Sam A. Golden ternalizing pathology that potentially functions through dys-
[email protected]
regulation of reward processing in a manner akin to drug
1 addiction (Blair 2016; Chester and DeWall 2016). This is
Department of Biological Structure, University of Washington,
Seattle, WA, USA supported by the common finding that, in humans, some neu-
2 ropsychiatric disorders present comorbid with inappropriate
Graduate Program in Neuroscience, University of Washington,
Seattle, WA, USA aggression, including depression and substance abuse (Tyrer
3 et al. 2015; Fazel et al. 2015), and recidivism rates of violent
Center of Excellence in Neurobiology of Addiction, Pain, and
Emotion (NAPE), University of Washington, Seattle, WA, USA offenders closely mimic relapse rates for drug-addicted
 Psychopharmacology

individuals (Sinha 2011; Ducrose et al. 2014). We and others we will briefly review the state of the appetitive aggression
have hypothesized that studying the motivational component literature (also see Golden and Shaham 2018; Golden et al.
of aggression, termed appetitive aggression, may provide nov- 2019b), and then within this context, provide a primer on how
el therapeutic approaches to the treatment of maladaptive ag- machine learning approaches (regarding both acquisition/
gression presenting both within and outside of neuropsychiat- tracking and predictive classifier analysis) may be incorporat-
ric comorbidities (Miczek et al. 2015, 2017; Golden et al. ed into future studies. Definitions of commonly used terms in
2019b; Covington et al. 2019). Unfortunately, from a preclin- computational neuroethology are included in Table 1. We
ical perspective, there are relatively few established models propose, predominantly thanks to efforts of numerous labs
for mechanistically studying the neurobiological basis of mal- in developing and advancing machine learning methods for
adaptive appetitive aggression (Hashikawa et al. 2018; behavioral tracking (He et al. 2017; Mathis et al. 2018;
Flanigan and Russo 2019; Golden et al. 2019b). Graving et al. 2019; Pereira et al. 2019), that currently avail-
Aggression motivation has long been a focus of preclinical able approaches are sufficient to overcome the main limita-
research (Thompson 1963), invariably examined through the tions preventing wide adoption of machine learning for scor-
lens of ethological analysis, but often stymied due to the tre- ing complex social behavior within the context of preclinical
mendous effort and durations required to manually score com- aggression research.
plex social behavior with sampling frequencies and accuracies
that match modern neuroscience techniques. Several recent
reviews have highlighted the power of machine learning ap- Appetitive aggression
proaches for creating automated behavioral classifiers to study
social behavior, including aggression (Anderson and Perona Winners like to win: revisiting aggression reward
2014; Egnor and Branson 2016; Robie et al. 2017; Gris et al.
2017; Brown and de Bivort 2018; Akay and Hess 2019; Sixty years of behavioral research have shown that the oppor-
Mathis and Mathis 2019; Datta et al. 2019). The ability to tunity for, or experience of, an aggressive encounter with a
recognize and categorize common behaviors in model species conspecific can be reinforcing in many species, including se-
is an integral component of model reproducibility and lect mammals. Early work established the propensity of
extendibility, and these observations can be combined with Siamese fighting fish (Beta splendens) to perform operant
computational neuroethology to circumvent several inherent tasks for the opportunity to attack a static (Thompson 1963;
issues with manual annotation—most notably including ob- Thompson and Sturm 1965) or animate representation of a
server drift and bias, long analysis times, and inter-rater reli- conspecific (Craft et al. 2003, 2007; Elcoro et al. 2008).
ability (Anderson and Perona 2014; Egnor and Branson 2016; Similar behaviors were also demonstrated in homing pigeons
Datta et al. 2019). Additionally, both supervised and unsuper- (Cole and Parker 1971) and male fighting cocks (Thompson
vised machine learning algorithms have uncovered previously 1964). Although this pioneering work showed that animals
unknown behavioral repertoires in model organisms and have will seek the opportunity for aggression (once thought to be
confirmed foundational assumptions of ethology (Vogelstein a uniquely human trait), these studies did not explore the re-
et al. 2014; Wiltschko et al. 2015; Rudolf et al. 2019). Several lationship between seeking aggression and the actual experi-
machine learning–based open-source packages have been de- ence of an aggressive encounter.
veloped that can track one or multiple animals during freely Subsequent work allowing physical attacks against conspe-
moving behavior, while others, including one recently re- cifics has highlighted the importance of dominance in solidi-
leased by our lab (Nilsson et al. 2020), can be used to classify fying aggression reward seeking. Among many species, those
behaviors based on pose estimation tracking (Table 2). These animals that win their first aggressive encounter are more like-
approaches have led to a renaissance in the study of ethology, ly to win subsequent bouts, while those that lose are more
which is poised to catapult classical behavioral neuroscience likely to lose subsequent bouts. Such “winner” and “loser”
into the realm of “big data.” effects (Ginsburg and Allee 1942; Oyegbile and Marler
However, even as these techniques have removed several 2005, 2006; Kudryavtseva et al. 2011), and have been dem-
hurdles impeding high-throughput behavioral analysis, the onstrated in meadow voles (Vlautin and Ferkin 2013),crayfish
use of automated behavioral classifiers has proven oddly dif- (Momohara et al. 2016), flour beetles (Okada et al. 2019),
ficult to adopt and generalize across labs. Implementation of Drosophila (Trannoy et al. 2015; Kim et al. 2018), lobster
these techniques is often slowed by a lack of computational cockroaches (Kou et al. 2019), Syrian hamsters (Schwartzer
knowledge, the need for specialized and expensive equip- et al. 2013), and many species and strains of mice (Oyegbile
ment, and the high computational expense to adequately train and Marler 2006). Some mammals do not display the winner
new classifiers. Due to the complexity of these programs, even effect, however, and species such as the white-footed mouse
the initial installation can be intimidating or difficult regard- only do so following experimental manipulations of testoster-
less of previous programming experience. In this perspective, one levels (Fuxjager et al. 2011). Aggressive encounters may
Psychopharmacology

Table 1 Definitions of common terms in computational neuroethology

Term Description

Pose The relationships and location of body part “key-points” (e.g., individual joints/digits in motor control tasks;
or the nose, ears, flanks, centroid, and tail base in an open-field task) that are relevant for determining an
individual’s activity in a video frame.
Pose estimation Computer-generated predictions of the individual's pose in a video frame.
Behavior_classification A computer-generated categorical prediction describing the presence or absence of a behavior
(e.g., attack versus no attack) in a video frame.
Features Independent variables derived from pose estimation (e.g., animal velocities and the distances between animals)
that are used to predict the presence and absence of a behavior in a video frame.
Feature weight/importance The relevance of a specific feature for accurately classifying the presence and absence of a behavior in a video
frame. For example, the distance between the first animal’s nose and the second animal’s tail base has
importance or weight for accurately classifying anogenital sniffing.
Ensemble algorithm The grouping of multiple learning algorithms based on different subsets of features. The individually weak
learning algorithms are combined to create a strong or accurate behavior classification (e.g., random forest algorithm).
Hidden states Unobservable factors determining the probability for ensuing expression of a behavior given the current behavior.
Model interpretability The extent to which a machine model’s decision processes can be understood by humans (e.g., how does the
algorithm discriminate between attack and non-attack events?).
Model explainability The extent to which a machine learning system allows a qualitative comprehension of the association between
the features of a specific observation and its prediction (e.g., why is the video frame classified as containing
an attack event?). Note that an explainable model also is an interpretable model.
“Black box” model A metaphor applied to machine learning algorithms with decision processes that are not readily explainable
or interpretable.

be sought prior to exposure, but winning often matters in Individual variability in inbred and outbred lines
maintaining the drive to seek aggression and in determining
future performance. A technical confound associated with preclinical aggression
To minimize the effects of bout outcome on behavioral research is the relative lack of innately expressed aggression
measures, many mouse assays for aggression use a varia- exhibited by commonly used inbred mouse strains relative to
tion on the resident-intruder task where a smaller subor- outbred strains (Jones and Brain 1987). Unlike outbred
dinate intruder is introduced into the home cage of a larg- strains, inbred mice often require significant experimenter ma-
er, older resident. Early studies using these procedures nipulation, either in the form of extended social isolation
found that mice will cross electric grids (Lagerspetz (Banerjee 1971) or repeated social instigation (Kudryavtseva
1964) and navigate T-mazes (Tellegen and Horn 1972; et al. 2014; Covington et al. 2018), to exhibit significant levels
Legrand 1978) and runways (Legrand 1970) to attack a of aggression towards conspecifics. To overcome this con-
subordinate. Furthermore, physical aggression alone is found, preclinical aggression research, and especially studies
sufficient to condition mice to prefer aggression-paired of appetitive aggression, has focused on the use of outbred
contexts (Potegal 1979; Taylor 1979). While these early mice (Chia et al. 2005) that exhibit a spectrum of innate ag-
influential studies demonstrate the utility of mouse gression behavior (Golden et al. 2016). While inbred mice are
models in the study of appetitive aggression, subsequent typically preferred for many research applications, recent
work shifted toward using resident-intruder (Miczek and meta-analysis shows that inbred strains do not have greater
O’Donnell 1978; Brain et al. 1981) or sensory contact trait stability than outbred mice, and data from outbred mice
(Kudryavtseva et al. 1991) assays where pairings are re- may be more generalizable across conditions and populations
peated and the subsequent interactions are recorded and (Tuttle et al. 2018).
scored for a variety of behavioral measures. However, We and others have used CFW or CD-1 outbred mice as
such procedures may be biased toward reactive aggression these strains display several naturally occurring aggression
due to their inescapable and involuntary nature phenotypes ranging from compulsive addictive–like aggres-
(Kudryavtseva et al. 2011, 2014). To overcome these lim- sion seeking to aggression avoidance (Golden et al. 2016,
itations, recent work has focused on behavioral proce- 2017b). Unfortunately, the use of outbred strains precludes
dures that examine both reactive and appetitive aggression the use of genetically defined Cre recombinase–based
using operant conditioning tasks, and this approach is in- methods. To overcome this, we have introduced a hybrid
strumental for understanding the neurobiological differ- breeding strategy using the F1 hybrid offspring derived from
ences underlying different types of aggression. an inbred strain of interest and an outbred CD-1 (Golden et al.
 Psychopharmacology

2017a; Aleyasin et al. 2018). Following behavioral phenotyp- aggression experiences can transform non-rewarding aggres-
ing for baseline reactive and appetitive aggression levels be- sive encounters into a rewarding experience. Second, aggres-
tween inbred and hybrid populations, as well as molecular sion CPP is a learned phenomenon that can be acquired even
phenotyping for appropriate transgene expression, this ap- by initially non-aggressive mice. Specifically, we exposed a
proach introduces genetic selectivity to innately aggressive large cohort of non-aggressors to 10 days of repeated resident-
mice populations. We propose that this approach, in combi- intruder testing and found that 50% transitioned to variable
nation with more traditional but time-consuming backcrossing aggressors and exhibited aggression CPP. Lastly, aggression
of Cre lines onto outbred strains or the impending develop- CPP is persistent, lasting several weeks following the final
ment of outbred CRISPR transgenic lines, will provide a condition session (Golden et al. 2017a).
strong foundation for aggression research moving forward. Notably, the portion of CD-1 mice that fail to show aggres-
sion is small and has not been the focus of study due to the
Unconditioned versus conditioned aggression difficulty in screening animals and filling group sizes.
Advances in automated behavioral tracking, however, may
The conditioned place preference (CPP) procedure has histor- be able to distinguish these phenotypes at an earlier age, alle-
ically been used to evaluate the rewarding effects of drugs and viating these restrictions, as will be discussed in the second
alcohol (Beach 1957; Mucha et al. 1982). In this procedure, part of this review.
one distinct context is paired with the conditioned stimulus
while another context is paired with the unconditioned stimu- Addiction-like aggression behavior and relapse
lus. During a subsequent stimulus-free test, the laboratory an-
imal chooses to spend time within the conditioned or uncon- Beyond measures of reactive aggression and conditioned ag-
ditioned context. An increase in preference for the conditioned gression reward, several groups have developed operant tasks
stimulus paired context is indicative of rewarding effects that measure appetitive aggression. The Miczek group has
(Bardo and Bevins 2000). More recently, CPP procedures designed an operant conditioning panel including active and
have been developed to assess the relative reward of affiliative inactive nose poke ports that can be introduced into the home
social interactions (Panksepp and Lahvis 2007; Dölen et al. cage of outbred CFW mice. The resident mice are trained to
2013; Goodwin et al. 2018). nose poke on fixed ratio and fixed interval schedules of rein-
Similarly, based on studies in female Syrian hamsters forcement to attack intruders (Fish et al. 2002, 2005; Bannai
(Meisel and Joppa 1994) and male outbred OF-1 mice et al. 2007). The development of operant aggression tasks has
(Martínez et al. 1995), we have adapted a CPP procedure in allowed the pharmacological decoupling of aggression
combination with the resident-intruder social defeat procedure seeking versus aggression consumption behaviors. The
(Miczek et al. 1982; Kudryavtseva et al. 1991; Golden et al. Miczek group reported that the GABAa positive allosteric
2011) to study aggression reward in CD-1 mice (Golden et al. modulator allopregnanolone increases operant response rates
2016). Using this method, we first categorized unconditioned at lower doses than are required for increased attack behaviors
reactive aggression in adult CD-1 male mice through repeated (Fish et al. 2002), but that the effects are inhibited by the rise
daily resident-intruder assays with adolescent submissive in corticosterone that are necessary for both operant
C57BL/6J intruder mice. Mice that attacked the intruders dur- responding and escalated aggression behaviors (Fish
ing these screening assays (70%) were termed aggressors, et al. 2005). While alcohol administration increases the
while mice that did not attack (30%) were termed non-aggres- motivation to fight, these effects are distinct from fight-
sors. We then evaluated conditioned aggression motivation ing performance and were not impacted by the antago-
using the aggression CPP assay. Mice that displayed aggres- nism of NMDA or AMPA receptors (Covington et al.
sion during the initial screening tests developed aggression 2018). 5-HT1b receptor agonism was found to decrease
CPP, while those that did not attack the intruders demonstrat- attack intensity without changing operant responding
ed conditioned place aversion (Golden et al. 2016). In a series (Bannai et al. 2007).
of follow-up experiments, we parametrically explored the ag- The Kennedy group subsequently replicated and extended
gression CPP phenomenon and observed several key findings. these results, finding reliable operant aggression self-
First, based on the observation that unconditioned reactive administration in mice under progressive ratio, differential
aggression falls along a continuum in CD-1 mice, we charac- reinforcement of low rate behavior, and variable ratio
terized individual differences in aggression CPP by testing a reinforcement schedules (Couppis and Kennedy 2008;
third phenotype, termed “variable aggressors,” composed of May and Kennedy 2009). Local nucleus accumbens do-
mice that performed inconsistently when repeatedly exposed pamine receptor 1 or 2 antagonism in Swiss Webster
to the resident-intruder procedure (Golden et al. 2017a). The mice inhibited both operant responding and select attack
variable aggressive mice exhibited significant, although weak- behaviors (Couppis and Kennedy 2008), and we have
er, aggression CPP, suggesting that repeated unconditioned found similar results following chemogenetic inhibition
Psychopharmacology

of dopamine receptor (Drd) type 1, but not type 2, in 19%) that exhibited intense operant-reinforced attack be-
hybrid F1 CD-1 × Drd1-Cre or Drd2-Cre mice (Golden havior, decreased likelihood to select an alternative pal-
et al. 2019a, 2019b). Extensions of these operant proce- atable food reward over aggression, heightened relapse
dures have also shown that animals rapidly cease ag- vulnerability and progressive ratio responding, and resil-
gression self-administration when confronted with a ience to punishment-induced suppression of aggression-
non-submissive intruder (Falkner et al. 2016), further reinforced operant responding.
highlighting the necessity of winning in promoting ag- Importantly, this study found that contingent punish-
gression reward. ment is effective for suppressing aggression-seeking be-
Pathological aggression in humans mimics cardinal fea- havior in the majority of aggressive mice, but not in
tures of drug addiction. Aggressive encounters are often those exhibiting compulsive aggression-seeking behav-
sought despite severe negative consequences, pathological iors, and ultimately fails to prevent spontaneous recov-
aggression develops only in a minority of individuals ery of aggression-seeking following extended abstinence
(Lacourse et al. 2002; Provencal et al. 2015), and recidi- in nearly all aggressive mice. These data are especially
vism rates for violent offenders who are incarcerated for interesting in light of reports that in rodents, footshock
repeat violent offenses are similar to the relapse rates of elicits both unconditioned aggression (O’kelly and
individuals who take addictive drugs (Anthony et al. 1994; Steckle 1939; Azrin et al. 1967) and Pavlovian condi-
Ducrose et al. 2014). Such cardinal features have been tioned aggression to a footshock-paired tone (Vernon
reverse-translated from the clinic to create animal models and Ulrich 1966). However, work in non-human pri-
of addiction-like behavior. Deroche-Gamonet (2004) used mates (Ulrich et al. 1969; Azrin 1970) and rats
a combination of operant procedures including fixed and (Baenninger and Grossman 1969; Roberts and Blase
progressive ratio tasks, as well as fixed ratio with cue- 1971) have shown that mechanical pain or shock-
contingent shock punishment, within a rodent model of induced aggression is suppressed by contingent, but
cocaine addiction that has high face validity to the DSM not non-contingent, shock punishment. Further, non-
IV criteria (Deroche-Gamonet 2004). In a cohort of rats contingent footshock inhibits aggression in dominant
that initially showed equal levels of cocaine self- but not subordinate mice (Frischknecht et al. 1985).
administration and sensitization, 17% of animals went on Together, and within the common context of “punish-
to show addiction-like indicators including difficulty stop- ment as a tool to prevent aggression,” these data sug-
ping or limiting intake, high motivation for access, and gest greater nuance and present areas for future study
continued use despite negative consequence. These mea- within the context of adaptive and maladaptive compul-
sures also correlated with relapse propensity (Deroche- sive aggressive behavior.
Gamonet 2004; Piazza and Deroche-Gamonet 2013).
Based on the above considerations, we have developed a Conclusions
modified operant chamber to test relapse to aggression-
seeking following (i) home cage forced abstinence The recent renaissance in aggression research has been
(Pickens et al. 2011), (ii) voluntary choice-based absti- assisted by the integration of carefully designed behavior-
nence (Caprioli et al. 2015), and (iii) punishment-induced al assays, derived from animal models traditionally used
abstinence (Krasnova et al. 2014; Marchant et al. 2019). to study compulsive drug use and relapse in the addiction
Using this approach, we have shown that preclinical addic- field, and modern neuroscience techniques such as
tion models can be used to identify the neural mechanisms chemogenetics (Coward et al. 1998; Armbruster et al.
controlling appetitive aggression and relapse, as well as path- 2007) and optogenetics (Boyden et al. 2005). These ma-
ological or compulsive manifestations of aggression (Golden nipulations have begun to help us understand the neural
et al. 2017b). About 70% of aggressive mice learn to lever- regions driving attack behavior and aggression salience
press for aggressive interactions, and using several gold-stan- (Miczek et al. 2001; Falkner et al. 2016; Han et al.
dard models derived from the preclinical addiction liter- 2017; Stagkourakis et al. 2018; Golden et al. 2019a),
ature, we observed aggression relapse after forced absti- but the timescale of neural manipulations and recordings
nence, punishment-induced abstinence, or choice-based necessitate higher resolution behavioral scoring than hand
voluntary abstinence that persists long after the last ag- scoring is able to consistently and reliably provide. Both
gressive act. Through cluster analysis of the aggression- aggression-seeking and consumption phases of behavioral
related measures, we also identified a subset of mice assays will benefit from more rapid, reliable, and non-
that met criteria previously developed to denote compul- subjective quantification of behavior. Within the field of
sive addiction in rodent models (Deroche-Gamonet pharmacology, real-time quantification and behavioral
2004). Specifically, the cluster analysis identified a sub- prediction are necessary for closed-loop manipulations
set of compulsive addiction-like aggressive mice (~ that can causally differentiate the behavioral sequence
 Psychopharmacology

constituents of reactive and appetitive aggressive behav- Rudolf et al. 2019; Zhang and Yartsev 2019).
ior. Specific to current operant procedures, such ap- Markerless pose estimation and behavioral classification
proaches will allow a detailed understanding of the se- algorithms are rapidly improving, and many of the req-
quela of behavioral events that occur after an operant ac- uisites are met for expanding machine learning ap-
tion is contingently reinforced with an intruder. All of proaches to classify complex social behavior. A key
these time-locked events, at previously unfeasible time component to the adoption and success of these efforts
scales and objectivity, may provide additional information is promoting open-source packages that can be easily
on the motivational state of the aggressor and the neural and widely adopted throughout the behavioral neurosci-
mechanisms guiding this behavior. ence community.
Currently, aggression consumption outcome measures are There are several requirements for the ready adoption of
typically restricted to latency to first attack, or proportion of machine learning approaches:
attack versus non-attack behaviors. Such simplifications leave
valuable ethological data unexplored, but these analytical (1) Ease of use. The field of computational neuroethology is
omissions have been necessary due to the complexity of ag- advancing at rapid pace, such that the ability of many
gression behavior and the length of time and training required behavioral neuroscientists to take advantage of these
to accurately manually score assays. Species typical attack pipelines is curtailed by their skill behind a command
behavior in mice generally includes bites directed toward the line. Many of the recent additions to the field such as
back and flanks of intruder mice, and decreased attack behav- DeepLabCut (Mathis et al. 2018), however, provide
ior upon displays of submission by an intruder. Additionally, graphic user interfaces (GUIs) that allow the user to
shifts toward more damaging bite locations or continued at- avoid setting hyperparameters or organizing projects
tack despite submission can both be indicative of escalated via command line entries. Programs with easy installa-
aggression (Takahashi and Miczek 2014; Newman et al. tion and approachable GUIs will open doors for labs
2018). Automated behavioral classifiers can simultaneously without easy access to computationally inclined
measure these and other behaviors, of both the resident and individuals.
intruder animal, allowing for a richer ethological continuum. (2) Generalizability. Behavioral assays within and be-
The classifiers—once validated—can be curated and dissem- tween labs can often be filmed under different lighting
inated through online repositories, which can eliminate ob- conditions, with animals and backgrounds shifting in
server variability and present significant opportunities for hues, using variable hardware and acquisition param-
cross-site standards within behavioral analyses. Accessible eters such as frame rate, resolution, color scales, and
classifiers that are fast, operational across labs,and interpret- video format. Many current implementations of ma-
able and explainable (Table 1) also have obvious benefits for chine learning for animal behavior (Table 2) require
transparency and scientific rigor. specialized hardware setups that can be prohibitively
We and others propose that open-source machine learning expensive or challenging to implement and scale-up
techniques will allow for rapid, high-throughput explorations for high-throughput behavioral assays, and often do
of the incredible nuance of these behaviors without sacrificing not generalize well from one preparation to another
accuracy. The integration of these automated behavioral clas- (Anderson and Perona 2014). Machine learning ap-
sifiers will overcome the hurdle of hand scoring which has proaches that can use standard, readily accessible re-
bottlenecked the field since its emergence. cording hardware without the need for specialized
builds are vital for generalizability and comprehensive
implementation.
Machine learning (3) Cost management. Current pose estimation and behav-
ioral classifier programs benefit from the use of depth
Embracing machine learning cameras, multiple filming angles, specialized acqui-
sition hardware, and proprietary software. Access to
Automated video assessments can exceed human performance high-end analytical platforms, either in the form of
(Gris et al. 2017), and behavioral classifiers increase throughput cloud-based or local solutions, is a requirement for
and consistency (Schaefer and Claridge-Chang 2012) in addition generating both tracking and classifier models.
to reducing human bias and anthropomorphism in scor- These costs propel computational neuroethology
ing (Robie et al. 2017). The development of specialized out of the reach of many behavioral neuroscience
machine learning behavioral classifiers have uncovered labs. However, the behavioral assays commonly
previously unknown behavioral repertoires in animals utilized for the study of aggressive behavior are
including mice, drosophila, bats, and Caenorhabditis well defined and often recorded with a single cam-
elegans (Vogelstein et al. 2014; Wiltschko et al. 2015; era, making it possible to generate and share
Table 2 Currently available open-source tracking or behavioral classification software

Name Validated species Multiple animals? Main output Behavioral classifiers validated Mechanism Software website
in original publication

Ctrax (Branson et al. 2009) Drosophila Yes Behavior Walk, stop, sharp turn, crabwalk, Computer vision http://ctrax.sourceforge.net/
Psychopharmacology

backup, touch, chase

CADABRA (Dankert et al. 2009) Drosophila Yes Behavior Lunging, tussling, wing threat/extension, Supervised http://www.vision.caltech.
circling, chasing, copulation edu/cadabra/
Autotyping (Jhuang et al. 2010) Mice No Behavior Home-cage behaviors: drink, eat, groom, Supervised http://cbcl.mit.edu/software-
hang, micromovement, rear, rest, walk datasets/mouse/
JAABA (Kabra et al. 2013) Drosophila, mice Yes Behavior Flies: walk, stop, crabwalk, backup, touch, Supervised https://sourceforge.net/projects/
chase, jump, copulation, wing jaaba/files/
flick/grooming/extension, righting,
center pivot, tail pivot. Mice: follow,
walk.
MiceProfiler (de Chaumont Mice Yes, if no obstruction Behavior Head-head contact, anogenital contact, Physics engine http://icy.bioimageanalysis.
et al. 2012) side by side, approach, leave, follow, org/plugin/mice-
chase profiler-tracker/
Unnamed* (Burgos-Artizzu Mice Yes Behavior Approach, walk away, circle, chase, attack, Supervised
et al. 2012) copulation, drink, eat, clean, sniff, rear
Unnamed* (Matsumoto Rats Yes Behavior Rearing, head-head and head-hip contact, Physics engine See supplementary material
et al. 2013) approach, leave, follow, mount,
intromission, ejaculation
MotionMapper (Berman Flies Yes Behavior Grooming: wing, leg, abdomen, head. Unsupervised https://github.com/gordonberman/
et al. 2014) Wing waggle, running MotionMapper
Unnamed* (Hong et al. 2015) Mice If distinct Behavior Attack, close investigation, mounting Supervised
MoSeq* (Wiltschko et al. 2015) Mice No Behavior Dart, micromovement, pause, Unsupervised MTA and private repository
rear, walk
Unnamed (Unger et al. 2017) Mice Yes Behavior Direct contact, nose to nose, Unsupervised
anogenital sniffing, following,
mating, self-grooming
DuoMouse (Arakawa et al. 2017) Mice Yes Behavior Sniffing, following, indifferent Unsupervised http://www.mgrl-lab.jp/
DuoMouse.html
B-SOiD (Hsu and Yttri 2019) Mice No Behavior Pause, rear, groom, sniff, orient left/right, Unsupervised https://github.com/
locomotion YttriLab/B-SOID
SimBA (Nilsson et al. 2020) Mice, rats Yes Behavior Attack, pursuit, lateral threat, anogenital Supervised https://github.com/
sniff, sgoldenlab/simba
mounting, upright submissive,
allogrooming,
flee, scramble, boxing, approach,
avoidance,
drink, clean, eat, rear, walk away, circle
Stytra (Stih et al. 2019) Zebrafish larvae No Closed-loop tracking and Computer vision https://github.com/
stimulus admin. (real time) portugueslab/stytra
ToxID (Rodriguez et al. 2017; Fish, ants, mice Yes Individual tracking in Computer vision https://sourceforge.net/
also see ToxTrac groups (real time) projects/toxtrac/
Rodriguez et al. 2018)
MoST (Thanos et al. 2017) Mice Yes Individual tracking in groups Computer vision
idtracker.ai (Romero-Ferrero Drosophila, fish, ants, mice Yes Individual tracking in groups Unsupervised http://idtracker.ai/
et al. 2019, also see idTracker
Pérez-Escudero et al. 2014)
LocoMouse* (Machado et al. 2015) Mice No Kinematics Supervised https://github.com/careylab
/LocoMouse
DeepLabCut (Mathis et al. 2018) Drosophila, mice, Yes Pose Supervised http://www.mousemotorlab.
horses, humans, fish org/deeplabcut
 Psychopharmacology

“base” tracking and classifier models that preclude

https://github.com/talmo/leap

NeLy-EPFL/DeepFly3D
https://github.com/jgraving/
the need for individual labs to heavily invest in

https://sourceforge.net/
specialized hardware or commercial software.

projects/sos-track/
Software website

murthylab/sleap
(4) Accuracy. At a minimum, automated classifiers need to
https://github.com/

https://github.com/

DeepPoseKit
meet the accuracy of manual annotation conducted by
highly trained observers. Regardless of the adoption of
machine learning techniques, the ability to consistently
recognize behaviors of interest in model species is inte-
Computer vision gral to behavioral research, and the experimenter apti-
Mechanism

Unsupervised

tude to recognize significant behavioral events deter-

Supervised

Supervised

mines the success in machine model generation, tuning,

and validation. We have found, however, that even ob-
servers trained to high rates of inter-rater reliability on
strict operational definitions of behaviors (Table 3) may
occasionally miss shorter behavioral bouts when scoring
Behavioral classifiers validated

long video recordings (e.g., 20k+ video frames).

Machine learning approaches provide the advantage of
in original publication

never tiring, and in our experience, classifiers consistent-

ly detect shorter behavioral bouts missed by human ob-
servers. While ideal algorithms for specific tasks are typ-
ically dependent on data structure and the computational
knowledge within a lab, we and other labs have found
that supervised approaches (Nilsson et al. 2020), such as
random forest machine learning algorithms, provide an
appropriate balance of ease of use, interpretability,
Posture and trajectory (real

explainability, and accuracy that often exceeds human

annotation accuracy.
(5) Easy expansion of training sets. Ethological datasets pro-
Pose (real time)
Main output

vide an interesting challenge for machine learning.

time)

Training data provided to algorithms must be represen-

Pose

Pose

tative of the testing data, but behavioral assays are often

conducted in discrete cohorts with unique properties.
Multiple animals?

Yes, through SLEAP

Training sets should contain a balanced mix of control

and treatment videos, and diverse and unstable environ-
*Next to software name indicates multiple cameras or specialized equipment necessary

mental influences (cage change schedules, season-

If distinct

ally noisy HVAC systems, etc.) must be considered

No

No

as potential confounds if not represented in the

original classifier training data. Specific to pharma-
Drosophila, locusts, zebras

cological research, when individual animal move-

Drosophila, mice, fish
Validated species

ment characteristics and social behaviors may be

Drosophila, humans
Drosophila, mice

altered by pharmacological interventions, it is not feasi-

ble to create representative training sets from the outset.
We and others have approached this challenge by creat-
ing the ability to easily add additional hand annotated
videos (several short videos are typically sufficient) to
Estimates Animal Pose (SLEAP)

DeepPoseKit (Graving et al. 2019)

training sets, which can then rapidly be used to create

DeepFly3d* (Günel et al. 2019)

(Gomez-Marin et al. 2012)

updated, iterative, behavior classifiers, and animal track-

Sensory Orientation Software
Successor: Social LEAP
LEAP (Pereira et al. 2019)
Table 2 (continued)

ing models.
(6) Interpretability. Interpretable methods are of utmost
importance for reproducibility, transparency, and
rigor (Table 1). Methods for creating accurate clas-
Name

sification algorithms in behavioral neuroscience

should provide transparent processes allowing for
Psychopharmacology

Table 3 Operational definitions for aggression behaviors exhibited between resident CD-1 mice and intruder C57B6/J mice
Classifier Description Start frame Duration of behavior End frame
Attack Clear physical antagonistic First frame when the Resident mouse Short breaks (equal or less than First frame when Resident
interaction initiated by the makes physical antagonistic contact with 133ms, or 4 frames at 30fps) may be mouse orients away from
Resident mouse. Brief the Intruder. Typically, this is present in attack behavior, as long as Intruder. Typically, this is a
pauses can occur (equal or characterized by outstretched Resident Resident is still oriented toward slight turning of the head to
less than 133ms, or 4 forepaw(s) contacting the Intruder, while Intruder. Attacks can include tussling, look in a different direction,
frames at 30fps) as long as the Resident has an open mouth to initiate biting, boxing, and corralling as part followed by a relaxation of
the Resident mouse a bite. Can also be characterized by the of the attack bout. the body and moving away
remains oriented toward the first frame of a slap or quick bite without from the Intruder.
Intruder. the forepaws being outstretched.
Anogenital sniffing Resident mouse is sniffing First frame when the Resident mouse is Uninterrupted sniffing of anogenital First frame when Resident
the anogenital region of the clearly sniffing anogenital region of region. mouse moves head away
Intruder. Resident must be Intruder, rather than side, back, or leg. from anogenital region,
sniffing at base of tail, not either to move away from
further up on back or on Intruder or to sniff non-
legs. anogenital region.

Lateral threat Resident mouse is in close First frame when Resident mouse orients Resident will often circle Intruder or First frame when lateral
proximity (typically less than side to Intruder and tilts front of body move front half side to side in front of threat posture is dropped.
one body length away from toward Intruder. Intruder, feigning attacks prior to Animal will shift head away
the Intruder) to face of actual attack. to look at other target or
Intruder with back arched will begin an attack. If
and side displayed toward animal does not attack, the
Intruder. Ears are often back posture will relax.
pinned with shoulder and
side of face nearest to
Intruder mouse tilted slightly
toward ceiling.
Tail rattle Resident's tail is curved at Defined as rapid back and forth movement The tail continues to move rapidly First frame when the curled
two or more points (typically of tail. In Resident mice this manifests as back and forth and can occasionally side to side movement of
in an "S" shape) and curling of tail into "S" shape, where next straighten if tail is moving fast the tail stops. The tail
showing rapid back and frames show inversion of shape and rapid enough. Tail side to side movement typically relaxes and
forth movement. whipping back and forth of the tail. is continuous. straightens out.

Pursuit Resident mouse is following First frame when the Resident is moving The Intruder is specifically moving First frame when either
in the Intruders path as the toward Intruder as Intruder moves away. away from the Resident. The Intruder mouse stops moving, or
Intruder moves away from Typically, this is characterized by the is not sniffing the arena, foraging, the Resident deviates from
the Resident. etc. Intruder typically moves in a path of the Intruder.
and the Resident follows directly after, or straight line until it reaches the edge
when the Intruder walks past the Resident of the area, at which point it will turn
and the Resident markedly changes to face and watch the Resident.
directions to pursue the Intruder in the Resident is moving along the same
Intruders path. path as the Intruder without sniffing
or foraging in the arena.

clear understanding of how classifications are made These sequences can be further reduced to levels of granu-
and, ideally through GUI control, the capacity to larity which may or may not convey important information
titrate model parameters to operationalized stan- (e.g., face bite versus flank bite). Furthermore, on video,
dards. This may include the ability to visualize de- animals frequently occlude each other and significantly
cision paths and the importances of individual in- change their body shapes as they aggress, creating
dependent variables or features for the classification difficulties for pose estimators. To our knowledge, two
result, and how the data and features are balanced groups have published work using automated classifiers
within the model. These concepts have however— for evaluating aggression behaviors. Hong et al. (2015)
traditionally—not always been favored within com- was able to classify attack behaviors in male mice of differ-
puter and data sciences (Rudin 2019) and may in- ent colors using a multi-camera apparatus, and Burgos-
troduce additional challenges when creating acces- Artizzu et al. (2012) achieved accuracy rates of ~ 61% using
sible methods within preclinical social behavior and a multi-camera apparatus with two unmarked mice of the
aggression research. same color. Our lab has recently released Simple Behavioral
Analysis (SimBA) (Nilsson et al. 2020), an automated be-
Aggression behavioral assays provide excellent insight havioral classification technique for aggressive, submis-
into the current capabilities and limitations of machine sive, and social behaviors that was built upon pose estima-
learning algorithms due to the numerous technical chal- tion and the six principles outlined above.
lenges they present. Behavioral components often include The remainder of this review will discuss the current state
a very rapid succession of bites, lunges, pursuits, anogenital of machine learning for social behavior classification in ro-
sniffing, tail rattles, lateral threats, boxing, and occasionally dents in terms of aggression behavior, and the progress of the
head shakes and prosocial behaviors such as grooming. field toward meeting the goals discussed above.
 Psychopharmacology

Table 4 Recommended potential workflow and hardware specifications for performing supervised machine classification of complex social behaviors
in experimental animals using open-source software. For more information, see the GitHub repositories of the Golden lab

Workflow step Description Required Hardware Open-source tools

(i) Record videos at a quality where behaviors Camera allowing recordings at ≥ 30 fps FFmpeg
Record and of interest are apparent to trained and resolutions ≥ 640 × 480 OpenCV
pre-process human observers. (i.e., standard webcam).
videos Crop/clip video, downsample resolution,
and re-sample frame rates, if required.

(ii) Estimate the location of the animals and CPU supported by graphical processing DeepLabCut
Track animals their body parts in each frame unit (GPU), through a cloud solution DeepPoseKit
of the videos with (e.g., Google Colaboratory, AWS LEAP
convolutional neural network. SageMaker, YOLO
Microsoft Azure) or locally mask RCNN
(e.g., NVIDEA Rtx2080Ti).

(iii) Extract feature sets, annotate behaviors, Processing time benefits from higher-end scikit-learn
Supervised decision build/tune algorithms, evaluate classification CPUs and solid-state drives (SSDs). imblearn
ensembles performance. We recommend a CPU with ≥ 8 cores OpenCV
(e.g., Intel i9-9900K). Shapely

(iv) Visually evaluate machine classifications. Perform Processing time benefits from higher-end OpenCV
Visualize/analyze descriptive statistics of behavior sequences, their CPUs and solid-state drives (SSDs). FFmpeg
classification durations and patterns. We recommend a CPU with ≥ 8 cores matplotlib
results (e.g., Intel i9-9900K). standard python
library

Accurate behavior classifications require arenas and recording conditions that allow detailed animal tracking of relevant body parts through
convolutional neural network architectures (workflow step ii). Generally, if the recording conditions allow the behaviors of interest to be reliably
observed by human annotators in recorded videos, then the behaviors of interest can also be accurately annotated by machine classification techniques

Supervised versus unsupervised learning techniques can be tremendously powerful in identifying the
inherent structure present in behavior. Because the user does
Many machine learning techniques for behavioral analysis not predefine behaviors of interest, these algorithms have been
first identify the position of animals frame-by-frame in a video proposed to be less biased than supervised techniques. Thus
via pose estimation or background subtraction. In particular, far, unsupervised techniques have resulted in important ad-
accurate and rapid pose estimation via recently developed and vances in mouse ethology including the identification of facial
generalizable convolutional neural network architectures— expressions corresponding to neuronally separable emotion
accessible through packages such as DeepLabCut (Mathis states (Dolensek et al. 2020) and novel sub-second behavioral
et al. 2018), DeepPoseKit (Graving et al. 2019), and LEAP structures (Wiltschko et al. 2015). The program B-SOiD has
(Pereira et al. 2019)—provide a platform for generating the also combined unsupervised t-SNE clustering and support
pose estimation data needed to create rich feature sets required vector machines to great success in classifying naturally oc-
for machine classification of complex social behavior. For curring behaviors in single mice, within a pipeline that is read-
example, the predicted poses may be used to estimate dis- ily accessible to non-specialists (Hsu and Yttri 2019).The
tances between animals, their velocities, angles, and accelera- power to identify new behavioral repertoires is alluring
tions across rolling windows, and this may correlate with hu- and has the potential to significantly advance the field
man annotated instances of aggressive behavior and used in of ethology. Unsupervised machine learning tools are
further modeling techniques (Table 4). These positional data notoriously difficult to tune and interpret, however,
are then analyzed to cluster statistically similar images that the and we propose that they should not be a first-
program either identifies as a predefined behavior (supervised line option for generalization to labs without significant
learning) or the cluster is studied by the experimenter who computational experience.
then adds behavioral labels post hoc (unsupervised learning). Training an unsupervised behavioral classifier generally in-
Unsupervised learning has successfully classified behav- volves feeding many unlabeled video frames into a user-
iors in tunicates, drosophila, individual mice, and pairs of defined algorithm which then identifies and separates images into
mice after cropping the animals into individual videos behavioral clusters based on user-defined mathematical require-
(Vogelstein et al. 2014; Wiltschko et al. 2015; Klibaite et al. ments for differences and similarities. The de Bivort lab has
2017; Rudolf et al. 2019; Dolensek et al. 2020). These conducted one of the only large-scale direct comparisons of
Psychopharmacology

unsupervised learning techniques in classifying animal behavior computed labeling rules to future images. Pose estimators are
(the leg movement of flies) and provides an excellent overview powerful tools in identifying finer points of animal activity,
of unsupervised learning techniques (Todd et al. 2017). A typical and are able to identify phenotypes not detected with simply
pipeline for clustering often involves data pre-processing, dimen- trajectory tracking (Hong et al. 2015). Notably, new packages
sionality reduction, and cluster assignment (Todd et al. 2017; can include project management and image annotation GUIs,
Datta et al. 2019), though see (Wiltschko et al. 2015). which greatly enhances the user experience and has resulted in
Each of these steps is modular and can be performed using more widespread adoption.
any of several different algorithms. Pre-processing can include Supervised learning techniques can be used for both pose
transformations such as time frequency analysis, vector nor- estimation and behavioral classification, but unlike unsupervised
malization, and wavelet transformation, while dimensionality algorithms, can only classify predetermined behaviors and can-
reduction is often accomplished via techniques such as prin- not identity novel behavioral repertoires. Training accurate su-
cipal component analysis (PCA) or t-distributed stochastic pervised machine learning classifiers requires the careful, frame-
neighbor embedding (t-SNE). The final clustering algorithms, by-frame annotation of a behavior of interest. Human variability
such as k-means clustering or Gaussian mixture modeling, are and annotation mistakes, such as erroneous labeling of aggres-
used to group the data into clusters that researchers can apply sion events as non-aggression events, can generate significant
behavioral labels towards (Todd et al. 2017). The flexibility of noise that propagates to future machine analyses (Frenay and
unsupervised learning approaches at each step makes them Verleysen 2014). It is therefore important to have precise behav-
very powerful in adapting to different datasets, but tuning ioral operational definitions that encompass and exclude the be-
parameters requires careful consideration. havioral events and non-events, respectively, and we present
Importantly, unsupervised learning techniques do not have examples of how such precise operational definitions may look
built-in performance metrics. Tuning them to high perceived in Table 3. While such precise descriptions are not current stan-
accuracy (e.g., all user-defined attacks are also defined by the dard laboratory practice, and functionally translate to extra time
algorithm as an attack) is in essence training the model to meet hand-annotating videos, they can promote the introduction of
particular output benchmarks, i.e., creating a supervised algo- standardized cross-laboratory definitions through shareable clas-
rithm (Brown and de Bivort 2018). Some classes produced by sifier repositories that ultimately increase replicability as
an unsupervised approach may also be difficult for users to discussed previously.
define as recognizable behaviors. Todd et al. (2017) propose Several behavioral classifiers using supervised machine
evaluating networks via parameters such as minimum dwell learning techniques have been validated in mice, including
time within a behavior class and the reliability of the algorithm JAABA (Kabra et al. 2013), autotyping (Jhuang et al. 2010),
when trained multiple times on the same data set (Todd et al. and SimBA (Nilsson et al. 2020). Each uses simple, single
2017; Brown and de Bivort 2018). These parameters acknowl- camera setups with significant differences in underlying ma-
edge the real-world constraints on behavior and the need for chine learning approaches to create predictive classifiers.
replication while imposing less bias onto the model in terms of JAABA classifies walking and following behavior in groups
expected output. We propose that any implementation of al- of same colored mice using a boosting ensemble algorithm
gorithmic features during unsupervised training should be ac- (Kabra et al. 2013), and autotyping uses a combination of
companied by a description of the experimental objectives hidden Markov model and support vector machines to clas-
driving their necessity. sify a variety of home cage behaviors in single-housed mice
Supervised learning algorithms require greater user over- (Jhuang et al. 2010). SimBA uses random forest classifiers to
sight in that users are required to annotate training data that the classify aggressive, defensive, and other social behaviors in
classification system can use to learn to correctly label images. mice following pose estimation (Nilsson et al. 2020). Some
Several pose estimators are validated for use in drosophila programs also use multiple cameras, depth cameras, or other
(Branson 2014; Berman et al. 2014; Günel et al. 2019), and specialized setups for behavioral classification. The
tools such as DeepLabCut (Mathis et al. 2018), LEAP (Pereira Anderson Lab has developed a random forest classification
et al. 2019), and DeepPoseKit (Graving et al. 2019) have been technique which can identify attack, mounting, and investi-
validated in mice and other non-human mammals. gation behaviors in differently colored mice using multiple
DeepLabCut and LEAP use variations on the supervised depth sensing cameras (Hong et al. 2015). Boosting ensemble
learning technique of deep neural networks to estimate animal algorithms have also been used to classify social movement,
pose, while the DeepPoseKit uses a novel Stacked attack, copulation, exploratory, and drinking and eating be-
DenseNet algorithm. For in-depth explanations of these esti- haviors with integrated top and side view filming (Burgos-
mators, see Mathis et al. (2018), Graving et al. (2019), Pereira Artizzu et al. 2012).
et al. (2019), Nath et al. (2019), and Mathis and Mathis Whether implementing supervised or unsupervised ma-
(2019). In many of these packages, users label body parts at chine learning techniques, there is a strong argument for using
set coordinates, and these training labels are used to generalize algorithms that are easily interpretable (Rudin 2019). Black
 Psychopharmacology

box models have been helpful in image analysis and interpretability and being both easy to tune and robust
uncovering new behavioral patterns, but the inability to read- against overfitting. Here, we describe several of these
ily understand the underlying assumptions of algorithms is a algorithms and highlight the potential pros and cons
detriment. Black box models—not unlike prevalent behavior- they possess regarding the classification of aggression-
al protocols—can appear to be valid and accurate (strong face related behaviors.
validity) while measuring factors unrelated to the phenome-
non it claims to measure (weak construct validity). Examples Neural networks (Hopfield 1982; LeCun et al. 1989)
of such fallacies, with more trivial consequences, would in-
clude convolutional neural networks that “successfully” dis- Neural networks or convoluted neural networks are typically
criminate husky dogs and wolfs based predominantly on used for large datasets with a high number (tens of thousands)
only the presence or absence of snow in the image of features and observations, as well as noise. Features may
(Ribeiro et al. 2016), or networks generating “success- include the trajectory of individual body parts, Euclidean dis-
ful” gender classifications from iris texture images tances, and body part movement over a small frame of time.
through the presence or absence of mascara Users provide an input layer of data to the neural network and
(Kuehlkamp et al. 2017). More seriously, black box define the desired categories in the output layer. The network
models are notorious for aggregated racial and social- then computes additional hidden layers between the input and
economic biases (Mittelstadt et al. 2016; Rudin 2019; output layers to correctly classify input data into the appropri-
Obermeyer et al. 2019). While clearly a different class ate output categories or classifiers. Hidden layers are advan-
of concern than what is anticipated in preclinical behav- tageous as specific features are not user-selected for inclusion
ioral computational neuroethology, these cautionary tales re- within the algorithm, but the cost is an absence of information
main relevant and important. Explainable and interpretable on what the resulting classifications are ultimately based on.
models should thus be adopted as best scientific practice Neural networks can be challenging to interpret and
(Rudin 2019) to allow for user oversight of the parameters tune, and similar performance can be achieved with oth-
guiding decision processes. er methods (Lietman et al. 1999; Nitze et al. 2012; Liu
While supervised learning techniques require users to pre- et al. 2013). Many programs, including the pose estima-
define behaviors of interest, they are often much easier to tune tors described above, successfully use deep neural net-
and interpret than unsupervised learning techniques, particu- works for pose and/or behavioral classifications
larly for labs without statisticians or computer scientists on (Karpathy et al. 2014; Krizhevsky et al. 2017; Mathis
staff. In the spirit of ease of use, generalizability, and accura- et al. 2018; Pereira et al. 2019). Due to the lack of
cy, we propose that supervised learning techniques are a good information regarding feature weights, we propose that
starting point for the automated examination of aggressive more transparent algorithms which perform at similar
behaviors. levels are preferable for behavioral classification derived
from tracked features.
Common classifying algorithms for supervised
learning Random forests (Breiman 2001; Liaw and Wiener 2002)

Within supervised learning, there are many different algo- Random forests are a type of ensemble algorithm. Ensemble
rithms from which to choose, and no method is universally algorithms are composed of many independently trained weak
superior to the others (Hand 2006). The appropriate selection models which are combined to make strong predictions.
depends on the structure, noise, and biases within each Random forests rely on the bootstrapping of both (i) subsets
dataset, as well as the threshold for acceptable training of data and (ii) predefined features to make a powerful forest
duration and willingness to troubleshoot parameters of decision trees than can then vote on the classification of a
(Hand 2006; Anderson and Perona 2014; Egnor and behavior. For example, a decision tree in a random forest starts
Branson 2016; Gris et al. 2017; Akay and Hess 2019). with a subset of data (e.g., 500 frames of 50,000) then uses a
Common techniques for behavioral and image classifi- subset of the features (e.g., 10 out of 100 features) to split the
cation include variations of neural networks, support data into yes/no classifications. Random forests are useful
vector machines, gradient boosting machines, random because each branch is created using the feature which pro-
forests, and hidden Markov models. All of these algo- vides the most information, and by combining these data from
rithms have potential benefits and may outperform the forest, it is possible to determine which features were most
others on specific datasets, but we and others have important for creating the behavioral classification. This tech-
found that random forest classifiers provide high accu- nique is computationally rapid as the trees are independent
racy classification (Breiman 2001; Liaw and Wiener and can be built in parallel, and they are unlikely to overfit
2002; Nilsson et al. 2020) with the added benefit of data based on noise (Breiman 2001).
Psychopharmacology

One weakness of random forests is their inability to native- (Caruana and Niculescu-Mizil 2006; Caruana et al. 2008).
ly support biased datasets. These datasets are common in be- Due to the larger number of tuning parameters involved in
havioral videos, in which most frames do not contain the be- support vector machines and the similar or superior classifica-
havior of interest. For example, a 5-min video filmed at 80 fps tion performance by random forest classifiers, we suggest that
with 10 total seconds of tail rattle would contain 23,200 random forests may be a better starting point for creating gen-
frames with no tail rattle and only 800 frames with tail rattle. eralizable and easily adopted classifiers.
Constructing a random forest with these data as is would
mostly construct trees with no or very few instances of tail
Hidden Markov models (Rabiner and Juang 1986)
rattle frames. In order to train the random forests more robust-
ly, over-sampling and/or under-sampling techniques (Chawla
Hidden Markov models (HMMs) can be integrated with su-
et al. 2002; Batuwita and Palade 2013) can be used to balance
pervised or unsupervised techniques and excel at finding pat-
the data.
terns in small sequences of data. These models have recently
been extended for use in behavioral classifications due to the
Gradient boosting machines (Freund and Schapire 1997;
similarities from frame to frame of video and the inherent
Friedman et al. 2000; Friedman 2001)
structure of behavior (Carola et al. 2011; Wiltschko et al.
2015; Arakawa et al. 2017). HMMs use hidden weights to
Gradient boosting machines build decision trees iteratively,
predict the probability of a transition from one state to another.
attempting to fix erroneous classifications after each node split
For example, if a resident is biting the flank of an intruder in
by finding a different feature which better classifies incorrect-
one frame, it is unlikely to be grooming its face in the next
ly classified data. Gradient boosting can be more robust than
frame. As indicated by the name of the algorithm, behavioral
random forests if tuned correctly but are more prone to
classifications depend on a hidden state. Anderson and Perona
overfitting data and incorrectly identifying noise, such as pose
(2014) raise an interesting point that HMMs may be valuable
misestimation, as legitimate behavioral data. While pose esti-
in objectively measuring behaviors resulting from the “emo-
mation is often highly accurate, there are outliers that may
tion state” of an animal. HMMs have been used in rodent
cause overfitting when fed into a gradient boosting machine.
aggression studies to investigate the effects of context (intrud-
With more accurate pose estimation and highly consistent
er behavior) on the aggression behavior of a home cage resi-
filming conditions across labs, gradient boosting machines
dent to parse species typical versus escalated aggression phe-
may provide superior classification than random forests, but
notypes (Haccou et al. 1988; Natarajan et al. 2009), and may
it is difficult to currently meet these conditions and maintain
provide an interesting path forward for understanding contex-
generality of video recording conditions. While the use of
tual aggression. While the hidden states can potentially be
gradient boosting machines is a good goal, their widespread
found via maximum a posteriori state estimations
adoption may depend on how well “base” models, and their
(Allahverdyan and Galstyan 2009), we propose that more in-
underlying recording acquisition parameters, are standardized
herently interpretable algorithms are preferable for initial be-
across laboratory recording environments.
havioral classification.
Support vector machines (Cortes and Vapnik 1995)
Promising directions
Support vector machines are frequently used and work best with
small datasets containing few outliers (Hsu and Lin 2002; Xu We strongly reiterate that there is no single correct pipeline for
2006). Using multiple tuning parameters (kernel, regularization, machine learning classification of animal behavior, but rather
gamma, and margin), support vector machines identify clusters that the most successful approaches will likely use multiple
of similar and dissimilar data in multidimensional space and find pipelines in parallel while taking advantage of their pros and
the regression line that best separates clusters based on the train- attempting to diminish their cons. The powerful techniques
ing data. Understanding linear algebra and principle component discussed above are summarized in Table 5, and the choice
analysis, in addition to having appreciable patience with tuning between them often revolves around the applicable knowl-
parameters, is helpful when using support vector machines. Like edge within a lab and which algorithm best classifies data
random forests, support vector machines do not work well na- during pilot testing (Table 5). To facilitate generalizability to
tively with biased datasets, which can similarly be overcome new labs, however, random forests seem like an excellent
through re-balancing and over- and under-sampling techniques. option due to their relatively few input parameters, robustness
There are few comparative studies of classification algo- to noise, relative interpretability and explainability, and high
rithms due to lack of generalization to other datasets, but in performance on a wide array of data. A future goal of auto-
those that exist, support vector machines do not typically per- mated behavioral analysis is long-term real-time tracking of
form as well as random forests or gradient boosting machines individuals within groups in a naturalistic setting, and many
 Psychopharmacology

Table 5 Overview of advantages and limitations of select open-source options for tracking and behavioral classification

Technique Open-source Advantages Limitations

alternative(s)

Segmentation/identification idtracker.ai • Can track large groups • Requires GPU support

networks for of unmarked individuals • No live tracking
“fingerprinting” • Accessible GUI • Only provides trajectory of animal
centroid(s)
• Not validated in mouse/rat video
with high
amounts of aggression
• Users must generate their social behavior
classifiers from tracking data, which may
also require additional layer of pose
estimation processing.

RFID/depth camera system Live Mouse Tracker • Built-in battery of social behavior • Manufacture cost (~ $2 k)
classifiers • Manufacturing time
• Group analysis of unmarked • Dedicated platform (cannot be used
animals (≤ 4 mice) to analyze
• No model training required historical datasets)
• Live tracking (30 Hz) • Arena/RFID implementation require
• Accessible SQL data format Animal Study Protocol
• True longitudinal analysis (days) submissions/revision
• Validated for mice only

Pose estimation DeepLabCut • Accurate tracking of • CNN training time

LEAP user-defined body parts • Requires GPU access
DeepPoseKit • Very large user-base • Group and live tracking features not yet
• Active code development available
• Accessible GUI’s (LEAP/DLC) • Users generate their social behavior
• Accepts most recording classifiers from tracking data
environments/ and video formats

Behavioral classification SimBA • Accessible GUI’s •Requires third-party tracking (and GPU sup-
JAABA • User-defined social behavior classi- port)
fiers • No live classifications
• Can accept any tracking data as input • Social classification restricted by
the inability of
third-party tracking solutions to track multiple
similarly coat-colored animals

The techniques listed above are not mutually exclusive, and the most successful behavioral analyses likely depend on synergistic implementations of
disparate computational methods

groups are currently working with these and other techniques program LEAP (sLEAP) and DeepLabCut have recently been
to solve parts of this problem. released, and provide pose estimation for multiple interacting
Distinguishing similarly looking animals has historically animals (Mathis et al. 2018; Pereira et al. 2019) by taking
been a challenge for automated pose estimators, and ultimate- advantage of algorithms such as optical flow (Brox et al.
ly problematic for the computational study of social behavior 2006), part affinity fields (Cao et al. 2017), and deep learning.
as many experiments require the use of nearly identical The program idTracker uses separate segmentation and iden-
looking animals. Although not a problem when using outbred tification networks to individually identify animals, and their
CD-1 mice as residents and C57 mice as intruders, due to their trajectories and back propagation techniques to identify ani-
difference in coat color, this does preclude the generality of mals during partial occlusions (Pérez-Escudero et al. 2014).
machine learning approaches to many rodent social behavioral Beyond image detection, there are several groups working to
assays. There are several recent programs that work to track integrate RFID tracking and video tracking (Weissbrod et al.
the identity of individual and similarly appearing animals in 2013; de Chaumont et al. 2019; Peleh et al. 2019). Live Mouse
groups over time, including ToxID, MoST, and idTracker.ai Tracker (de Chaumont et al. 2019) uses a combination of RFID
(Thanos et al. 2017; Rodriguez et al. 2017; Romero-Ferrero tracking and depth imaging, and has been validated with groups
et al. 2019). Additionally, social tracking extensions of the of up to 4 mice in a semi-natural environment over time. These
Psychopharmacology

programs are highly influential in allowing for the long-term Conclusion

study of group dynamics in large, enriched home cage environ-
ments and currently provide the state-of-art tracking solutions Aggression is an immutable force that contributes to the suf-
for mice of similar appearance. A commercial version of this fering and death for millions of people around the world
program, RFID-Assisted SocialScan, is also available and inte- (Sumner et al. 2015). While aggression can be highly reward-
grates a large behavioral arena with attached nest boxes (Peleh ing and pursued despite adverse consequences (Chester and
et al. 2019). Together, these programs can track the identity of DeWall 2016; Gan et al. 2019), as well as sought after lengthy
individuals while examining social hierarchies over long pe- abstinence (Ducrose et al. 2014), aggression is not generally
riods of time without excessive manual oversight or behavioral viewed in the medical profession or by the public as a com-
scoring. However, these approaches require significant invest- ponent of neuropsychiatric disease states (Golden et al. 2017a;
ments in highly specialized hardware. Golden and Shaham 2018). Indeed, in the recent formulation
Several groups pioneering pose estimation are also contin- of Research Domain Criteria (RDoC) of NIMH, the term ag-
ually working to increase the speed of their algorithms to gression does not appear under any of the research domains.
achieve real-time results (Graving et al. 2019; Mathis and Not surprisingly, no progress has been made in the treatment
Mathis 2019). Currently, ToxID and Sensory Orientation of pathological aggression, and little progress has been made
Software are able to provide real-time tracking of individual using aggression as a diagnostic tool for neuropsychiatric co-
mice and their trajectories (Gomez-Marin et al. 2012; morbidity (Martin et al. 2013). We believe that drawing greater
Rodriguez et al. 2017), and DeepLabCut has recently been attention to the neurobiological mechanisms of maladaptive
used for real-time pose estimation (Forys et al. 2018). appetitive aggression within the context of neuropsychiatric
Extending these platforms for in-depth real-time behavior- disease will both act to (i) de-stigmatize aggression comorbidity
al classification in mice will allow for the evaluation of in a manner similar to current approaches to substance abuse
behavior on the timescale of neural activity and allow fully treatments and (ii) support the identification of novel mechanis-
“closed-loop” recording and manipulation studies. The tic therapeutic approaches, which have presently been relegated
Stytra package is currently validated in zebrafish larvae, to neuroleptic dopamine antagonists like haloperidol (Ostinelli
and is able to both analyze pose and activity and rapidly et al. 2017) that exert their clinical efficacy through the neuro-
provide stimuli to the experimental animal based on its leptic’s sedative effects (Calver et al. 2015). Therefore, we pro-
activity (Stih et al. 2019). pose that neurobiological and behavioral tools used to study
These open-source packages are continuing to improve and drug seeking and relapse should be used to study brain mech-
emerge, and integration between them is becoming easier. anisms of appetitive aggression, both preclinically and clinical-
Pose estimators that can rapidly track individual animals de- ly. Further, expanding beyond reactive aggression to incorpo-
spite multiple occlusions is an essential first step for the auto- rate the full spectrum of ethologically relevant aggressive be-
mation of aggression assays, and our lab has focused heavily haviors is an essential step in increasing the utility and transla-
on exploring random forest classifiers for the classification of tion of ongoing preclinical research. Machine learning tools
aggression behaviors. Integrating the pose estimators above enabling high-throughput automated behavioral analysis are
with high accuracy and easy to use classifiers in an open- key to pursuing these lines of inquiry at a pace and depth nec-
source GUI allows for the generalization of these techniques essary for modern neuroscience methods.
across labs while maintaining the ability to take advantage of Significant progress has been made in developing open-
newer tracking technologies as they emerge. source software packages that are capable of tracking social
For example, our lab has recently developed an open- behavior over long periods of time and across diverse acqui-
source graphical workflow (for more information, see the sition parameters (Table 5), and we propose that a focus on
SimBA GitHub repository) for creating supervised decision increasing ease of use, generalizability, cost-consciousness,
ensembles from body part tracking data generated through accuracy, and easy expansion of training sets will allow for
convolutional neural networks in accordance with the consid- their wider adoption. Such automated scoring promises to
erations and principles discussed above (Nilsson et al. 2020). remove long-standing bottlenecks within aggression research
The package incorporates accessible menus for pre-processing and allow for the high-throughput experiments required for
and annotating videos, generating/deploying classification al- mapping the nuance of aggression phenotypes while simulta-
gorithms, with advanced machine learning validation, tuning neously building a common language of aggression pheno-
and visualization tools. The models created through this types through shared machine models across labs.
software—together with extensive documentation and
tutorials—are available for the scientific community through Acknowledgments We thank Briana Smith, Liana Bloom, Sophia
Hwang, and Jia Jie Choong for their thoughtful assistance in discussing
online repositories, and we hope that this can spur further data
the manuscript.
sharing efforts that boosts the accuracy and scope of machine
classification techniques in preclinical aggression research.
 Psychopharmacology

F u n d i n g i n fo r m at io n Th e re se arch wa s s up po rt ed NIDA Berman GJ, Choi DM, Bialek W, Shaevitz JW (2014) Mapping the ste-
4R00DA045662-02 (SAG), NIDA P30 DA048736 (SRON and SAG)), reotyped behaviour of freely moving fruit flies. J R Soc Interface 11:
NARSAD Young Investigator Award 27082 (SAG), and NIDA T32 20140672. https://doi.org/10.1098/rsif.2014.0672
5T32NS099578-04 (NLG). Blair RJR (2016) The neurobiology of impulsive aggression. J Child
Adolesc Psychopharmacol 26:4–9. https://doi.org/10.1089/cap.
2015.0088
Compliance with ethical standards Boyden ES, Zhang F, Bamberg E, Nagel G, Deisseroth K (2005)
Millisecond-timescale, genetically targeted optical control of neural
Conflict of interest The authors declare that they have no conflict of activity. Nat Neurosci 8:1263–1268. https://doi.org/10.1038/
interest. nn1525
Brain PF, Benton D, Childs G, Parmigiani S (1981) The effect of the type
of opponent in tests of murine aggression. Behav Process 6:319–
327. https://doi.org/10.1016/0376-6357(81)90049-8
References Branson K (2014) Distinguishing seemingly indistinguishable animals
with computer vision. Nat Methods 11:721–722. https://doi.org/
10.1038/nmeth.3004
Akay A, Hess H (2019) Deep learning: current and emerging applications Branson K, Robie AA, Bender J, et al (2009) High-throughput ethomics
in medicine and technology. IEEE J Biomed Health Inform 23:906– in large groups of Drosophila. Nat Methods 6:451–457. https://doi.
920. https://doi.org/10.1109/JBHI.2019.2894713 org/10.1038/nmeth.1328
Aleyasin H, Flanigan ME, Golden SA, Takahashi A, Menard C, Pfau Breiman L (2001) Random forests. Mach Learn 45:5–32. https://doi.org/
ML, Multer J, Pina J, McCabe KA, Bhatti N, Hodes GE, 10.1023/A:1010933404324
Heshmati M, Neve RL, Nestler EJ, Heller EA, Russo SJ (2018) Brown AEX, de Bivort B (2018) Ethology as a physical science. Nat
Cell-type-specific role of ΔFosB in nucleus accumbens in modulat- Phys 14:653–657. https://doi.org/10.1038/s41567-018-0093-0
ing intermale aggression. J Neurosci 38:5913–5924. https://doi.org/ Brox T, Rosenhahn B, Cremers D, Seidel H-P (2006) High accuracy
10.1523/JNEUROSCI.0296-18.2018 optical flow serves 3-D pose tracking: exploiting contour and flow
Allahverdyan A, Galstyan A (2009) On maximum a posteriori estimation based constraints. In: Leonardis A, Bischof H, Pinz A (eds)
of hidden Markov processes. arXiv:09061980 [cond-mat, physics: Computer vision – ECCV 2006. Springer, Berlin, Heidelberg, pp
physics, stat] 98–111
Anderson DJ, Perona P (2014) Toward a science of computational ethol- Burgos-Artizzu XP, Dollar P, Dayu Lin, et al (2012) Social behavior
ogy. Neuron 84:18–31. https://doi.org/10.1016/j.neuron.2014.09. recognition in continuous video. In: 2012 IEEE Conference on
005 Computer Vision and Pattern Recognition. IEEE, Providence, RI,
Anthony JC, Warner LA, Kessler RC (1994) Comparative epidemiology pp 1322–1329
of dependence on tobacco , alcohol , controlled substances , and Calver L, Drinkwater V, Gupta R et al (2015) Droperidol v. haloperidol
inhalants : basic findings from the national comorbidity survey for sedation of aggressive behaviour in acute mental health:
Arakawa T, Tanave A, Takahashi A et al (2017) Automated estimation of randomised controlled trial. Br J Psychiatry 206:223–228. https://
mouse social behaviors based on a hidden Markov model. In: doi.org/10.1192/bjp.bp.114.150227
Westhead DR, Vijayabaskar MS (eds) Hidden Markov models: Cao Z, Simon T, Wei S-E, Sheikh Y (2017) Realtime multi-person 2D
methods and protocols. Springer New York, New York, NY, pp pose estimation using part affinity fields. arXiv:161108050 [cs]
185–197 Caprioli D, Venniro M, Zeric T, Li X, Adhikary S, Madangopal R,
Armbruster BN, Li X, Pausch MH et al (2007) Evolving the lock to fit the Marchant NJ, Lucantonio F, Schoenbaum G, Bossert JM, Shaham
key to create a family of G protein-coupled receptors potently acti- Y (2015) Effect of the novel positive allosteric modulator of metab-
vated by an inert ligand. Proc Natl Acad Sci U S A 104:5163–5168. otropic glutamate receptor 2 AZD8529 on incubation of metham-
https://doi.org/10.1073/pnas.0700293104 phetamine craving after prolonged voluntary abstinence in a rat
Azrin NH (1970) Punishment of elicited aggression. J Exp Anal Behav model. Biol Psychiatry 78:463–473. https://doi.org/10.1016/j.
14:7–10. https://doi.org/10.1901/jeab.1970.14-7 biopsych.2015.02.018
Azrin NH, Hutchinson RR, Hake DF (1967) Attack, avoidance, and Carola V, Mirabeau O, Gross CT (2011) Hidden Markov model analysis
escape reactions to aversive shock. J Exp Anal Behav 10:131–148. of maternal behavior patterns in inbred and reciprocal hybrid mice.
https://doi.org/10.1901/jeab.1967.10-131 PLoS One 6:e14753. https://doi.org/10.1371/journal.pone.0014753
Caruana R, Niculescu-Mizil A (2006) An empirical comparison of super-
Baenninger R, Grossman JC (1969) Some effects of punishment on pain-
vised learning algorithms. In: Proceedings of the 23rd international
elicited aggression. J Exp Anal Behav 12:1017–1022. https://doi.
conference on machine learning - ICML ’06. ACM Press,
org/10.1901/jeab.1969.12-1017
Pittsburgh, Pennsylvania, pp 161–168
Banerjee U (1971) An inquiry into the genesis of aggression in mice Caruana R, Karampatziakis N, Yessenalina A (2008) An empirical eval-
induced by isolation. Behaviour 40:86–99 uation of supervised learning in high dimensions. In: Proceedings of
Bannai M, Fish EW, Faccidomo S, Miczek KA (2007) Anti-aggressive the 25th international conference on machine learning - ICML ’08.
effects of agonists at 5-HT1B receptors in the dorsal raphe nucleus ACM Press, Helsinki, Finland, pp 96–103
of mice. Psychopharmacology 193:295–304. https://doi.org/10. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE:
1007/s00213-007-0780-5 Synthetic minority over-sampling technique. J Artif Intell Res 16:
Bardo MT, Bevins RA (2000) Conditioned place preference: what does it 321–357. https://doi.org/10.1613/jair.953
add to our preclinical understanding of drug reward? Chester DS, DeWall CN (2016) The pleasure of revenge: retaliatory
Psychopharmacology 153:31–43 aggression arises from a neural imbalance toward reward. Soc
Batuwita R, Palade V (2013) Class imbalance learning methods for sup- Cogn Affect Neurosci 11:1173–1182. https://doi.org/10.1093/scan/
port vector machines. In: He H, Ma Y (eds) Imbalanced learning. nsv082
John Wiley & Sons, Inc., Hoboken, NJ, USA, pp 83–99 Chia R, Achilli F, Festing MFW, Fisher EMC (2005) The origins and
Beach HD (1957) Effect of morphine on the exploratory drive. Can J uses of mouse outbred stocks. Nat Genet 37:1181–1186. https://doi.
Psychol 11:237–244 org/10.1038/ng1665
Psychopharmacology

Cole JM, Parker BK (1971) Schedule-induced aggression: access to an Fish EW, De Bold JF, Miczek KA (2002) Aggressive behavior as a
attackable target bird as a positive reinforcer. Psychon Sci 22:33–35. reinforcer in mice: activation by allopregnanolone.
https://doi.org/10.3758/BF03335924 Psychopharmacology 163:459–466. https://doi.org/10.1007/
Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20: s00213-002-1211-2
273–297. https://doi.org/10.1007/BF00994018 Fish EW, DeBold JF, Miczek KA (2005) Escalated aggression as a re-
Couppis MH, Kennedy CH (2008) The rewarding effect of aggression is ward: corticosterone and GABAA receptor positive modulators in
reduced by nucleus accumbens dopamine receptor antagonism in mice. Psychopharmacology 182:116–127. https://doi.org/10.1007/
mice. Psychopharmacology 197:449–456. https://doi.org/10.1007/ s00213-005-0064-x
s00213-007-1054-y Flanigan ME, Russo SJ (2019) Recent advances in the study of aggres-
Covington HE, Newman EL, Tran S et al (2018) The urge to fight: sion. Neuropsychopharmacol 44:241–244. https://doi.org/10.1038/
persistent escalation by alcohol and role of NMDA receptors in s41386-018-0226-2
mice. Front Behav Neurosci 12:206. https://doi.org/10.3389/fnbeh. Forys B, Xiao D, Gupta P et al (2018) Real-time markerless video track-
2018.00206 ing of body parts in mice using deep neural networks. bioRxiv:
Covington HE, Newman EL, Leonard MZ, Miczek KA (2019) 482349. https://doi.org/10.1101/482349
Translational models of adaptive and excessive fighting: an emerg- Frenay B, Verleysen M (2014) Classification in the Presence of Label
ing role for neural circuits in pathological aggression. F1000Res 8. Noise: A Survey. IEEE Trans Neural Netw Learning Syst 25:845–
https://doi.org/10.12688/f1000research.18883.1 869. https://doi.org/10.1109/TNNLS.2013.2292894
Coward P, Wada HG, Falk MS, Chan SDH, Meng F, Akil H, Conklin BR Freund Y, Schapire RE (1997) A decision-theoretic generalization of on-
(1998) Controlling signaling with a specifically designed Gi- line learning and an application to boosting. J Comput Syst Sci 55:
coupled receptor. Proc Natl Acad Sci U S A 95:352–357 119–139. https://doi.org/10.1006/jcss.1997.1504
Craft BB, Velkey AJ, Szalda-Petree A (2003) Instrumental conditioning Friedman JH (2001) Greedy function approximation: a gradient boosting
of choice behavior in male Siamese fighting fish (Betta splendens). machine. Ann Stat 29:1189–1232
Behav Process 63:171–175 Friedman J, Hastie T, Tibshirani R (2000) Additive logistic regression: a
Craft BB, Szalda-Petree AD, Brinegar JL, Haddad NF (2007) Effect of statistical view of boosting (with discussion and a rejoinder by the
various discriminative stimuli on choice behavior in male Siamese authors). Ann Stat 28:337–407. https://doi.org/10.1214/aos/
fighting fish (Betta splendens). Percept Mot Skills 104:575–580. 1016218223
https://doi.org/10.2466/pms.104.2.575-580 Frischknecht H-R, Siegfried B, Waser PG (1985) Postaggression
Dankert H, Wang L, Hoopfer ED, Anderson DJ, Perona P (2009) footshock inhibits aggressive behavior in dominant but not in iso-
Automated monitoring and analysis of social behavior in lated mice. Behav Neural Biol 44:132–138. https://doi.org/10.1016/
Drosophila. Nat Methods 6:297–303. https://doi.org/10.1038/ S0163-1047(85)91286-5
nmeth.1310 Fuxjager MJ, Montgomery JL, Marler CA (2011) Species differences in
Datta SR, Anderson DJ, Branson K et al (2019) Computational the winner effect disappear in response to post-victory testosterone
neuroethology: a call to action. Neuron 104:11–24. https://doi.org/ manipulations. Proc Biol Sci 278:3497–3503. https://doi.org/10.
10.1016/j.neuron.2019.09.038 1098/rspb.2011.0301
de Chaumont F, Ey E, Torquet N, Lagache T, Dallongeville S, Imbert A, Gan G, Zilverstand A, Parvaz MA, Preston-Campbell RN, d’Oleire
Legou T, le Sourd AM, Faure P, Bourgeron T, Olivo-Marin JC Uquillas F, Moeller SJ, Tomasi D, Goldstein RZ, Alia-Klein N
(2019) Real-time analysis of the behaviour of groups of mice via a (2019) Habenula-prefrontal resting-state connectivity in reactive ag-
depth-sensing camera and machine learning. Nat Biomed Eng 3: gressive men – a pilot study. Neuropharmacology 156:107396.
930–942. https://doi.org/10.1038/s41551-019-0396-1 https://doi.org/10.1016/j.neuropharm.2018.10.025
Deroche-Gamonet V (2004) Evidence for addiction-like behavior in the Ginsburg B, Allee WC (1942) Some effects of conditioning on social
rat. Science 305:1014–1017. https://doi.org/10.1126/science. dominance and subordination in inbred strains of mice. Physiol
1099020 Zool 15:485–506
Dölen G, Darvishzadeh A, Huang KW, Malenka RC (2013) Social re- Golden SA, Shaham Y (2018) Aggression addiction and relapse: a new
ward requires coordinated activity of nucleus accumbens oxytocin frontier in psychiatry. Neuropsychopharmacol 43:224–225. https://
and serotonin. Nature 501:179–184. https://doi.org/10.1038/ doi.org/10.1038/npp.2017.173
nature12518 Golden SA, Covington HE, Berton O, Russo SJ (2011) A standardized
Dolensek N, Gehrlach DA, Klein AS, Gogolla N (2020) Facial expres- protocol for repeated social defeat stress in mice. Nat Protoc 6:1183–
sions of emotion states and their neuronal correlates in mice. Science 1191. https://doi.org/10.1038/nprot.2011.361
368:89–94. https://doi.org/10.1126/science.aaz9468 Golden SA, Heshmati M, Flanigan M, Christoffel DJ, Guise K, Pfau ML,
Ducrose MR, Alexia D. Cooper, Howard N. Snyder (2014) Recidivism of Aleyasin H, Menard C, Zhang H, Hodes GE, Bregman D, Khibnik
prisoners released in 30 states in 2005: patterns from 2005 to 2010. L, Tai J, Rebusi N, Krawitz B, Chaudhury D, Walsh JJ, Han MH,
U.S. Department of Justice Shapiro ML, Russo SJ (2016) Basal forebrain projections to the
Egnor SER, Branson K (2016) Computational analysis of behavior. Annu lateral habenula modulate aggression reward. Nature 534:688–692.
Rev Neurosci 39:217–236. https://doi.org/10.1146/annurev-neuro- https://doi.org/10.1038/nature18601
070815-013845 Golden SA, Aleyasin H, Heins R et al (2017a) Persistent conditioned
Elcoro M, Silva SP, Lattal KA (2008) Visual reinforcement in the female place preference to aggression experience in adult male sexually-
Siamese fighting fish, Betta splendens. J Exp Anal Behav 90:53–60. experienced CD-1 mice: persistent aggression conditioned place
https://doi.org/10.1901/jeab.2008.90-53 preference in CD-1 mice. Genes Brain Behav 16:44–55. https://
Falkner AL, Grosenick L, Davidson TJ, Deisseroth K, Lin D (2016) doi.org/10.1111/gbb.12310
Hypothalamic control of male aggression-seeking behavior. Nat Golden SA, Heins C, Venniro M, Caprioli D, Zhang M, Epstein DH,
Neurosci 19:596–604. https://doi.org/10.1038/nn.4264 Shaham Y (2017b) Compulsive addiction-like aggressive behavior
Fazel S, Wolf A, Chang Z, Larsson H, Goodwin GM, Lichtenstein P in mice. Biol Psychiatry 82:239–248. https://doi.org/10.1016/j.
(2015) Depression and violence: a Swedish population study. biopsych.2017.03.004
Lancet Psychiatry 2:224–232. https://doi.org/10.1016/S2215- Golden SA, Jin M, Heins C et al (2019a) Nucleus accumbens Drd1-
0366(14)00128-X expressing neurons control aggression self-administration and
 Psychopharmacology

aggression seeking in mice. J Neurosci 39:2482–2496. https://doi. melanogaster males. PNAS 115:1099–1104. https://doi.org/10.
org/10.1523/JNEUROSCI.2409-18.2019 1073/pnas.1716612115
Golden SA, Jin M, Shaham Y (2019b) Animal models of (or for) aggres- Klibaite U, Berman GJ, Cande J, Stern DL, Shaevitz JW (2017) An
sion reward, addiction, and relapse: behavior and circuits. J Neurosci unsupervised method for quantifying the behavior of paired animals.
39:3996–4008. https://doi.org/10.1523/JNEUROSCI.0151-19.2019 Phys Biol 14:015006. https://doi.org/10.1088/1478-3975/aa5c50
Gomez-Marin A, Partoune N, Stephens GJ, Louis M (2012) Automated Kou R, Hsu C-C, Chen S-C et al (2019) Winner and loser effects in
tracking of animal posture and movement during exploration and lobster cockroach contests for social dominance. Horm Behav
sensory orientation behaviors. PLoS One 7:e41642. https://doi.org/ 107:49–60. https://doi.org/10.1016/j.yhbeh.2018.12.002
10.1371/journal.pone.0041642 Krasnova IN, Marchant NJ, Ladenheim B, McCoy MT, Panlilio LV,
Goodwin NL, Lopez SA, Lee NS, Beery AK (2018) Comparative role of Bossert JM, Shaham Y, Cadet JL (2014) Incubation of methamphet-
reward in long-term peer and mate relationships in voles. Horm amine and palatable food craving after punishment-induced absti-
Behav 111:70–77. https://doi.org/10.1016/j.yhbeh.2018.10.012 nence. Neuropsychopharmacology 39:2008–2016. https://doi.org/
Graving JM, Chae D, Naik H, et al (2019) DeepPoseKit, a software 10.1038/npp.2014.50
toolkit for fast and robust animal pose estimation using deep learn- Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification
ing. 39 with deep convolutional neural networks. Commun ACM 60:84–
Gris KV, Coutu J-P, Gris D (2017) Supervised and unsupervised learning 90. https://doi.org/10.1145/3065386
technology in the study of rodent behavior. Front Behav Neurosci: Kudryavtseva NN, Bakshtanovskaya IV, Koryakina LA (1991) Social
11. https://doi.org/10.3389/fnbeh.2017.00141 model of depression in mice of C57BL/6J strain. Pharmacol
Günel S, Rhodin H, Morales D, et al (2019) DeepFly3D: a deep learning- Biochem Behav 38:315–320. https://doi.org/10.1016/0091-
based approach for 3D limb and appendage tracking in tethered, 3057(91)90284-9
adult Drosophila. bioRxiv 640375. https://doi.org/10.1101/640375 Kudryavtseva NN, Smagin DA, Bondar NP (2011) Modeling fighting
Haccou P, Kruk MR, Meelis E, van Bavel ET, Wouterse KM, Meelis W deprivation effect in mouse repeated aggression paradigm. Prog
(1988) Markov models for social interactions: analysis of electrical Neuro-Psychopharmacol Biol Psychiatry 35:1472–1478. https://
stimulation in the hypothalamic aggression area of rats. Anim Behav doi.org/10.1016/j.pnpbp.2010.10.013
36:1145–1163. https://doi.org/10.1016/S0003-3472(88)80074-5 Kudryavtseva NN, Smagin DA, Kovalenko IL, Vishnivetskaya GB
Han W, Tellez LA, Rangel MJ et al (2017) Integrated control of predatory (2014) Repeated positive fighting experience in male inbred mice.
hunting by the central nucleus of the amygdala. Cell 168:311– Nat Protoc 9:2705–2717. https://doi.org/10.1038/nprot.2014.156
324.e18. https://doi.org/10.1016/j.cell.2016.12.027 Kuehlkamp A, Becker B, Bowyer K (2017) Gender-from-iris or gender-
Hand DJ (2006) Classifier technology and the illusion of progress. Stat from-mascara? arXiv:170201304 [cs]
Sci 21:1–14. https://doi.org/10.1214/088342306000000060 Lacourse E, Côté S, Nagin DS et al (2002) A longitudinal–experimental
Hashikawa K, Hashikawa Y, Lischinsky J, Lin D (2018) The neural approach to testing theories of antisocial behavior development. Dev
mechanisms of sexually dimorphic aggressive behaviors. Trends Psychopathol 14:909–924. https://doi.org/10.1017/
Genet 34:755–776. https://doi.org/10.1016/j.tig.2018.07.001 S0954579402004121
He K, Gkioxari G, Dollár P, Girshick R (2017) Mask R-CNN. arXiv: Lagerspetz K (1964) Studies on the aggressive behaviour of mice.
170306870 [cs] Annales Academiae Scientiarum Fennicae Series B 131:1–131
Hong W, Kennedy A, Burgos-Artizzu XP, Zelikowsky M, Navonne SG, LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W,
Perona P, Anderson DJ (2015) Automated measurement of mouse Jackel LD (1989) Backpropagation applied to handwritten zip code
social behaviors using depth sensing, video tracking, and machine recognition. Neural Comput 1:541–551. https://doi.org/10.1162/
learning. Proc Natl Acad Sci U S A 112:E5351–E5360. https://doi. neco.1989.1.4.541
org/10.1073/pnas.1515982112 Legrand R (1970) Successful aggression as the reinforcer for runway
Hopfield JJ (1982) Neural networks and physical systems with emergent behavior of mice. Psychon Sci 20:303–305. https://doi.org/10.
collective computational abilities. PNAS 79:2554–2558. https://doi. 3758/BF03329080
org/10.1073/pnas.79.8.2554 Legrand R (1978) Reinforcing effect of aggressive behaviors preparatory
Hsu C-W, Lin C-J (2002) A comparison of methods for multiclass sup- to fighting in mice. Bull Psychon Soc 11:359–362. https://doi.org/
port vector machines. IEEE Trans Neural Netw 13:415–425. https:// 10.3758/BF03336854
doi.org/10.1109/72.991427 Liaw A, Wiener M (2002) Classification and regression by random.
Hsu AI, Yttri EA (2019) B-SOiD: an open source unsupervised algorithm Forest 2:6
for discovery of spontaneous behaviors. Neuroscience Lietman T, Eng J, Katz J, Quigley HA (1999) Neural networks for visual
Jhuang H, Garrote E, Yu X, Khilnani V, Poggio T, Steele AD, Serre T field analysis: how do they compare with other algorithms? J
(2010) Automated home-cage behavioural phenotyping of mice. Glaucoma 8:77–80
Nat Commun 1:1–10. https://doi.org/10.1038/ncomms1064 Liu M, Wang M, Wang J, Li D (2013) Comparison of random forest,
Jones SE, Brain PF (1987) Performances of inbred and outbred laboratory support vector machine and back propagation neural network for
mice in putative tests of aggression. Behav Genet 17:87–96. https:// electronic tongue data classification: application to the recognition
doi.org/10.1007/BF01066013 of orange beverage and Chinese vinegar. Sensors Actuators B Chem
Kabra M, Robie AA, Rivera-Alba M, Branson S, Branson K (2013) 177:970–980. https://doi.org/10.1016/j.snb.2012.11.071
JAABA: interactive machine learning for automatic annotation of Marchant NJ, Campbell EJ, Pelloux Y, Bossert JM, Shaham Y (2019)
animal behavior. Nat Methods 10:64–67. https://doi.org/10.1038/ Context-induced relapse after extinction versus punishment: similar-
nmeth.2281 ities and differences. Psychopharmacology 236:439–448. https://
Karpathy A, Toderici G, Shetty S, et al (2014) Large-scale video classi- doi.org/10.1007/s00213-018-4929-1
fication with convolutional neural networks. In: 2014 IEEE Martin LA, Neighbors HW, Griffith DM (2013) The experience of symp-
Conference on Computer Vision and Pattern Recognition. IEEE, toms of depression in men vs women: analysis of the National
Columbus, OH, USA, pp 1725–1732 Comorbidity Survey Replication. JAMA Psychiatry 70:1100–
Kim Y-K, Saver M, Simon J, Kent CF, Shao L, Eddison M, Agrawal P, 1106. https://doi.org/10.1001/jamapsychiatry.2013.1985
Texada M, Truman JW, Heberlein U (2018) Repetitive aggressive Martínez M, Guillén-Salazar F, Salvador A, Simón VM (1995)
encounters generate a long-lasting internal state in Drosophila Successful intermale aggression and conditioned place preference
Psychopharmacology

in mice. Physiol Behav 58:323–328. https://doi.org/10.1016/0031- O’kelly LI, Steckle LC (1939) A note on long enduring emotional re-
9384(95)00061-m sponses in the rat. The Journal of Psychology 8:125–131. https://doi.
Mathis MW, Mathis A (2019) Deep learning tools for the measurement of org/10.1080/00223980.1939.9917655
animal behavior in neuroscience. arXiv:190913868 [cs, q-bio] Obermeyer Z, Powers B, Vogeli C, Mullainathan S (2019) Dissecting
Mathis A, Mamidanna P, Cury KM, Abe T, Murthy VN, Mathis MW, racial bias in an algorithm used to manage the health of populations.
Bethge M (2018) DeepLabCut: markerless pose estimation of user- Science 366:447–453. https://doi.org/10.1126/science.aax2342
defined body parts with deep learning. Nat Neurosci 21:1281–1289. Okada K, Okada Y, Dall SRX, Hosken DJ (2019) Loser-effect duration
https://doi.org/10.1038/s41593-018-0209-y evolves independently of fighting ability. Proc Biol Sci 286:
May ME, Kennedy CH (2009) Aggression as positive reinforcement in 20190582. https://doi.org/10.1098/rspb.2019.0582
mice under various ratio- and time-based reinforcement schedules. J Ostinelli EG, Brooke-Powney MJ, Li X, Adams CE (2017) Haloperidol
Exp Anal Behav 91:185–196. https://doi.org/10.1901/jeab.2009.91- for psychosis-induced aggression or agitation (rapid
185 tranquillisation). Cochrane Database Syst Rev 7:CD009377–
Meisel RL, Joppa MA (1994) Conditioned place preference in female CD009377. https://doi.org/10.1002/14651858.CD009377.pub3
hamsters following aggressive or sexual encounters. Physiol Oyegbile TO, Marler CA (2005) Winning fights elevates testosterone
Behav 56:1115–1118. https://doi.org/10.1016/0031-9384(94) levels in California mice and enhances future ability to win fights.
90352-2 Horm Behav 48:259–267. https://doi.org/10.1016/j.yhbeh.2005.04.
Miczek KA, O’Donnell JM (1978) Intruder-evoked aggression in isolated 007
and nonisolated mice: effects of psychomotor stimulants and L-do- Oyegbile TO, Marler CA (2006) Weak winner effect in a less aggressive
pa. Psychopharmacology 57:47–55. https://doi.org/10.1007/ mammal: correlations with corticosterone but not testosterone.
bf00426957 Physiol Behav 89:171–179. https://doi.org/10.1016/j.physbeh.
Miczek KA, Thompson ML, Shuster L (1982) Opioid-like analgesia in 2006.05.044
defeated mice. Science 215:1520–1522. https://doi.org/10.1126/ Panksepp JB, Lahvis GP (2007) Social reward among juvenile mice.
science.7199758 Genes Brain Behav 6:661–671. https://doi.org/10.1111/j.1601-
Miczek KA, Maxson SC, Fish EW, Faccidomo S (2001) Aggressive 183X.2006.00295.x
behavioral phenotypes in mice. Behav Brain Res 125:167–181. Peleh T, Bai X, Kas MJH, Hengerer B (2019) RFID-supported video
https://doi.org/10.1016/S0166-4328(01)00298-4 tracking for automated analysis of social behaviour in groups of
mice. J Neurosci Methods 325:108323. https://doi.org/10.1016/j.
Miczek KA, Takahashi A, Gobrogge KL, Hwa LS, de Almeida RMM
jneumeth.2019.108323
(2015) Escalated aggression in animal models: shedding new light
Pereira TD, Aldarondo DE, Willmore L, Kislin M, Wang SSH, Murthy
on mesocorticolimbic circuits. Curr Opin Behav Sci 3:90–95.
M, Shaevitz JW (2019) Fast animal pose estimation using deep
https://doi.org/10.1016/j.cobeha.2015.02.007
neural networks. Nat Methods 16:117–125. https://doi.org/10.
Miczek KA, DeBold JF, Gobrogge K, et al (2017) The role of neuro-
1038/s41592-018-0234-5
transmitters in violence and aggression. In: The Wiley Handbook of
Pérez-Escudero A, Vicente-Page J, Hinz RC et al (2014) idTracker: track-
Violence and Aggression. American Cancer Society, pp 1–13
ing individuals in a group by automatic identification of unmarked
Mittelstadt BD, Allo P, Taddeo M, Wachter S, Floridi L (2016) The ethics animals. Nat Methods 11:743–748. https://doi.org/10.1038/nmeth.
of algorithms: mapping the debate. Big Data & Society 3: 2994
205395171667967. https://doi.org/10.1177/2053951716679679 Piazza PV, Deroche-Gamonet V (2013) A multistep general theory of
Momohara Y, Minami H, Kanai A, Nagayama T (2016) Role of cAMP transition to addiction. Psychopharmacology 229:387–413. https://
signalling in winner and loser effects in crayfish agonistic encoun- doi.org/10.1007/s00213-013-3224-4
ters. Eur J Neurosci 44:1886–1895. https://doi.org/10.1111/ejn. Pickens CL, Airavaara M, Theberge F, Fanous S, Hope BT, Shaham Y
13259 (2011) Neurobiology of the incubation of drug craving. Trends
Mucha RF, van der Kooy D, O’Shaughnessy M, Bucenieks P (1982) Neurosci 34:411–420. https://doi.org/10.1016/j.tins.2011.06.001
Drug reinforcement studied by the use of place conditioning in rat. Potegal M (1979) The reinforcing value of several types of aggressive
Brain Res 243:91–105. https://doi.org/10.1016/0006-8993(82) behavior: a review. Aggress Behav 5:353–373. https://doi.org/10.
91123-4 1002/1098-2337(1979)5:4<353::AID-AB2480050404>3.0.CO;2-7
Natarajan D, de Vries H, Saaltink D-J, de Boer SF, Koolhaas JM (2009) Provencal N, Booij L, Tremblay RE (2015) The developmental origins of
Delineation of violence from functional aggression in mice: an etho- chronic physical aggression: biological pathways triggered by early
logical approach. Behav Genet 39:73–90. https://doi.org/10.1007/ life adversity. J Exp Biol 218:123–133. https://doi.org/10.1242/jeb.
s10519-008-9230-3 111401
Nath T, Mathis A, Chen AC, Patel A, Bethge M, Mathis MW (2019) Rabiner LR, Juang BH (1986) An introduction to hidden Markov
Using DeepLabCut for 3D markerless pose estimation across spe- Models. 12
cies and behaviors. Nat Protoc 14:2152–2176. https://doi.org/10. Ribeiro MT, Singh S, Guestrin C (2016) “Why Should I Trust You?”:
1038/s41596-019-0176-0 explaining the predictions of any classifier. In: Proceedings of the
Newman EL, Terunuma M, Wang TL, Hewage N, Bicakci MB, Moss SJ, 22nd ACM SIGKDD International Conference on Knowledge
DeBold JF, Miczek KA (2018) A role for prefrontal cortical NMDA Discovery and Data Mining - KDD ’16. ACM Press, San
receptors in murine alcohol-heightened aggression. Francisco, California, USA, pp 1135–1144
Neuropsychopharmacology 43:1224–1234. https://doi.org/10. Roberts CL, Blase K (1971) Elicitation and punishment of intraspecies
1038/npp.2017.253 aggression by the same stimulus. J Exp Anal Behav 15:193–196.
Nilsson SRO, Goodwin NL, Choong JJ et al (2020) Simple Behavioral https://doi.org/10.1901/jeab.1971.15-193
Analysis (SimBA): an open source toolkit for computer classifica- Robie AA, Seagraves KM, Egnor SER, Branson K (2017) Machine vi-
tion of complex social behaviors in experimental animals. bioRxiv sion methods for analyzing social interactions. J Exp Biol 220:25–
2020.04.19.049452. https://doi.org/10.1101/2020.04.19.049452 34. https://doi.org/10.1242/jeb.142281
Nitze I, Schulthess U, Asche H (2012) Comparison of machine learning Rodriguez A, Zhang H, Klaminder J, Brodin T, Andersson M (2017)
algorithms random forest, artificial neural network and support vec- ToxId: an efficient algorithm to solve occlusions when tracking
tor machine to maximum likelihood for supervised crop type classi- multiple animals. Sci Rep 7:1–8. https://doi.org/10.1038/s41598-
fication. 7 017-15104-2
 Psychopharmacology

Romero-Ferrero F, Bergomi MG, Hinz RC, Heras FJH, de Polavieja GG Thompson TI (1964) Visual reinforcement in fighting cocks. J Exp Anal
(2019) idtracker.ai: tracking all individuals in small or large collec- Behav 7:45–49. https://doi.org/10.1901/jeab.1964.7-45
tives of unmarked animals. Nat Methods 16:179–182. https://doi. Thompson T, Sturm T (1965) Visual-reinforcer color, and operant behav-
org/10.1038/s41592-018-0295-5 ior in Siamese fighting fish. J Exp Anal Behav 8:341–344. https://
Rudin C (2019) Stop explaining black box machine learning models for doi.org/10.1901/jeab.1965.8-341
high stakes decisions and use interpretable models instead. Nat Todd JG, Kain JS, de Bivort BL (2017) Systematic exploration of unsu-
Mach Intell 1:206–215. https://doi.org/10.1038/s42256-019-0048-x pervised methods for mapping behavior. Phys Biol 14:015002.
Rudolf J, Dondorp D, Canon L, Tieo S, Chatzigeorgiou M (2019) https://doi.org/10.1088/1478-3975/14/1/015002
Automated behavioural analysis reveals the basic behavioural rep- Trannoy S, Chowdhury B, Kravitz EA (2015) Handling alters aggression
ertoire of the urochordate Ciona intestinalis. Sci Rep 9:2416. https:// and “loser” effect formation in Drosophila melanogaster. Learn
doi.org/10.1038/s41598-019-38791-5 Mem 22:64–68. https://doi.org/10.1101/lm.036418.114
Schaefer AT, Claridge-Chang A (2012) The surveillance state of behav- Tuttle AH, Philip VM, Chesler EJ, Mogil JS (2018) Comparing pheno-
ioral automation. Curr Opin Neurobiol 22:170–176. https://doi.org/ typic variation between inbred and outbred mice. Nat Methods 15:
10.1016/j.conb.2011.11.004 994–996. https://doi.org/10.1038/s41592-018-0224-7
Schwartzer JJ, Ricci LA, Melloni RH (2013) Prior fighting experience Tyrer P, Kendall T, Barnett R, et al (2015) Violence and aggression:
increases aggression in Syrian hamsters: implications for a role of shortterm management in mental health, health and community
dopamine in the winner effect. Aggress Behav 39:290–300. https:// settings
doi.org/10.1002/ab.21476 Ulrich R, Wolfe M, Dulaney S (1969) Punishment of shock-induced
Sinha R (2011) New findings on biological factors predicting addiction aggression. J Exp Anal Behav 12:1009–1015. https://doi.org/10.
relapse vulnerability. Curr Psychiatry Rep 13:398–405. https://doi. 1901/jeab.1969.12-1009
org/10.1007/s11920-011-0224-0 Vernon W, Ulrich R (1966) Classical conditioning of pain-elicited ag-
Stagkourakis S, Spigolon G, Williams P, Protzmann J, Fisone G, gression. Science 152:668–669. https://doi.org/10.1126/science.
Broberger C (2018) A neural network for intermale aggression to 152.3722.668
establish social hierarchy. Nat Neurosci 21:834–842. https://doi.org/
Vlautin C, Ferkin M (2013) The outcome of a previous social interaction
10.1038/s41593-018-0153-x
with a same-sex conspecific affects the behavior of meadow voles,
Stih V, Petrucco L, Kist A, Portugues R (2019) Stytra: an open-source,
Microtus pennsylvanicus. ETHOLOGY 119:212–220
integrated system for stimulation, tracking and closed-loop behav-
Vogelstein JT, Park Y, Ohyama T, Kerr RA, Truman JW, Priebe CE,
ioral experiments. 19
Zlatic M (2014) Discovery of brainwide neural-behavioral maps via
Sumner SA, Mercy JA, Dahlberg LL, Hillis SD, Klevens J, Houry D
multiscale unsupervised structure learning. Science 344:386–392.
(2015) Violence in the United States: status, challenges, and oppor-
https://doi.org/10.1126/science.1250298
tunities. JAMA 314:478–488. https://doi.org/10.1001/jama.2015.
Weissbrod A, Shapiro A, Vasserman G et al (2013) Automated long-term
8371
tracking and social behavioural phenotyping of animal colonies
Takahashi A, Miczek KA (2014) Neurogenetics of aggressive behavior:
within a semi-natural environment. Nat Commun 4:2018. https://
studies in rodents. Curr Top Behav Neurosci 17:3–44. https://doi.
doi.org/10.1038/ncomms3018
org/10.1007/7854_2013_263
Taylor GT (1979) Reinforcement and intraspecific aggressive behavior. Wiltschko AB, Johnson MJ, Iurilli G, Peterson RE, Katon JM,
Behav Neural Biol 27:1–24. https://doi.org/10.1016/S0163- Pashkovski SL, Abraira VE, Adams RP, Datta SR (2015)
1047(79)92720-1 Mapping sub-second structure in mouse behavior. Neuron 88:
Tellegen A, Horn JM (1972) Primary aggressive motivation in three 1121–1135. https://doi.org/10.1016/j.neuron.2015.11.031
inbred strains of mice. J Comp Physiol Psychol 78:297–304. Xu L (2006) Robust support vector machine training via convex outlier
https://doi.org/10.1037/h0032192 ablation. 7
Thanos PK, Restif C, O’Rourke JR et al (2017) Mouse Social Interaction Zhang W, Yartsev MM (2019) Correlated neural activity across the brains
Test (MoST): a quantitative computer automated analysis of behav- of socially interacting bats. Cell 178:413–428.e22. https://doi.org/
ior. J Neural Transm 124:3–11. https://doi.org/10.1007/s00702-015- 10.1016/j.cell.2019.05.023
1487-0
Thompson TI (1963) Visual reinforcement in Siamese fighting fish. Publisher’s note Springer Nature remains neutral with regard to jurisdic-
Science 141:55–57. https://doi.org/10.1126/science.141.3575.55 tional claims in published maps and institutional affiliations.