Advanced Academic Writing

Week 6
11/9 (Wed)
Agenda: Discussion
• Assignment 1 & 2 Recap
• Discussion
• The purpose
• The structure
• Reverse-engineering
• Language Tips: Modal verbs
• Conclusions
• Reverse-engineering
• Assignment 5

2
Discussion!

4
 Structure of a research paper Title
Abstract
Title

Abstract Introduction
Introduction (Literature review)
Why?
(Literature
review)
Methods

Methods How? Results

Body
Discussion
Results What?
Conclusion
Discussion So what?
Acknowledgements
Conclusion
References References
The purpose of Discussion
• A narrative account that walks through the reader from Results to Conclusion
• To do this, identify the main value or contribution of your research to set the ‘destination’ that can be
clearly communicated to the reader via the narrative.
• Consider the following questions:
• Has my study obtained the same or similar results to previous studies but used a new that improved or
modified the existing method?
• The method may be your main contribution
• Has my study achieved better (such as more accurate) results than previous studies?
• The results themselves may be the main contribution
• Has my study obtained a game-changing result such that it sets a new direction for research or
invalidates previous research?
• The impact on the literature may be the main contribution
• Has my study identified or created new or extended applications
• The impact on industry or the real world may be the main contribution

• Prioritize the MAIN contribution

• Researchers rarely have a plenty of time reading a research article, so we must prioritize and
communicate the main contribution as clearly as possible.
6
Results vs. Discussion
• In the results section, you avoid making general conclusions,
speculations and comparison with other studies. Leave them for the
discussion part.
• Depending on the structure of your paper you may want to combine
results and discussion.
Results:
(What you did and) what were the results

Discussion:
What the results mean
7
 Generic model of Introduction
1. Establish the importance of the topic/field
2. Provide background information and the general problem area (or
research focus)
3. Present previous and/or current research and contributions
4. Locate a gap in the research and describe the problem you will address
5. Describe the present paper, sometimes mentioning aims, results,
methods, conclusions, and often including how the paper can
contribute to the field.
Discussion
 The field
Introduction Existing
vs. Discussion research
Gap
Your study

Revisit the study

Response to Gap
Map the study to
existing work
Your contribution

Adopted and modified from Glasman-Deal (2021). Science research writing for native
and nonnative speakers of English.
 Generic model of Discussion
1. Revisit some aspect of the study such as RQs, the gap in research, or the
key results
2. Discuss to what extent the study has responded to the gap or solved the
research problem
3. Position the study in relation to the research map by comparing the
results of the study to theory/model and previous research
4. Discuss the key takeaway and implication of the study to prepare the
reader to exit the article
Sample passage
Discussion and conclusions—In this Letter, we analyzed the relation between turbulent
heating rates and current sheets in a weakly collisional plasma by comparing directly both
quantities in real space. Using gyrokinetic simulations of KAW turbulence, we obtained
evidence that the locations of current sheets (defined as regions of high current) do not
generally correspond to peaks in the electron and ion heating rates. Several authors have
previously argued that heating in the solar wind occurs mainly in current sheets. However,
one needs to take into account that these analyses were performed either by using various
dissipation proxies instead of the actual local heating rates [7–10,13] or without measuring
directly the heating rate in real space [8].
Furthermore, we showed that plasma heating is highly anisotropic in velocity space,
providing first time direct measurements that identify the dominant collisional direction for
each species. We identified Landau damping as the preferred route for the electron
heating, while nonlinear (perpendicular) phase mixing is the channel responsible for ion
heating. These results demonstrate the importance of kinetic dynamics and the use of
well-defined collisional dissipation measures for studying plasma heating, further stressing
the need for in situ spacecraft measurements that would allow precise estimates of
collisional heating from particle distribution functions.
From Navarro, Alejandro Bañón, et al. "Structure of plasma heating in gyrokinetic Alfvénic 11
turbulence." Physical review letters 117.24 (2016): 245101.
Sample passage
Discussion and conclusions—In this Letter, we analyzed the relation between turbulent
heating rates and current sheets in a weakly collisional plasma by comparing directly both
quantities in real space. Using gyrokinetic simulations of KAW turbulence, we obtained
evidence that the locations of current sheets (defined as regions of high current) do not
generally correspond to peaks in the electron and ion heating rates. Several authors have
previously argued that heating in the solar wind occurs mainly in current sheets. However,
one needs to take into account that these analyses were performed either by using various
dissipation proxies instead of the actual local heating rates [7–10,13] or without measuring
directly the heating rate in real space [8].
Furthermore, we showed that plasma heating is highly anisotropic in velocity space,
providing first time direct measurements that identify the dominant collisional direction for
each species. We identified Landau damping as the preferred route for the electron
heating, while nonlinear (perpendicular) phase mixing is the channel responsible for ion
heating. These results demonstrate the importance of kinetic dynamics and the use of
well-defined collisional dissipation measures for studying plasma heating, further stressing
the need for in situ spacecraft measurements that would allow precise estimates of
collisional heating from particle distribution functions.
From Navarro, Alejandro Bañón, et al. "Structure of plasma heating in gyrokinetic Alfvénic 12
turbulence." Physical review letters 117.24 (2016): 245101.
 • State the achievement or the contribution of the study
• Revisit background information or literature to help the
reader
• Revisit the gap, aim, or method

Menu of • Review key results

• Map onto literature (both theory and empirical research)

Discussion
for comparison (support, not support)
• Identify limitations of the study and suggestions for future
research
• Restate the achievement and contribution and impact of
the study

There is no standard order for the components but

some of these tend to go together!
13
 • State the achievement or the contribution of the study
• Revisit background information or literature to help the
reader
• Revisit the gap, aim, or method

Menu of • Review key results

• Map onto literature (both theory and empirical research)
Discussion for comparison (support, not support)
• Identify limitations of the study and suggestions for future
research
• Restate the achievement and contribution and impact of
the study

14
Exercise: Recognizing components of Discussion
• Get in groups of 4-5 people
• Take 15 minutes to
• read the following two examples of Discussion
• discuss how each sentence or each group of
sentences functions within the section
• summarize your results of the discussion by coloring
each sentence or group of sentences using the
assigned color in the previous slide.
• summarize here:
https://docs.google.com/presentation/d/1yDko4Fg0E
ZaL9ArPUiIPyelWwm8hyp1iCfIqalw9zz4/edit?usp=sha
ring
• After this exercise, you will do the same for
your model article. 15
In this study, we fabricated a copper nanomesh and demonstrated its antimicrobial properties and
preservation of skin naturalness with conformality, thermal conductivity, and skin humidity. The antimicrobial
properties of copper were confirmed using the bacterial and viral inactivation tests, while further
enhancement of the large surface area of nanomesh was demonstrated. The prevention of cross-infection was
tested with a copper nanomesh by placing on the hand and interacting with various objects. The thin and
porous nanomesh structure provided ideal compatibility with the skin functions. In addition, the proposed
copper nanomesh exhibited advances in thermal transfer with a positive potential in thermal sensation and
object identification.
For the commercial application of the proposed copper nanomesh as an alternative to gloves or other skin
protective wear, there are a few issues remaining to be addressed. In this study, we chose thermally
evaporated copper as the antimicrobial material, but it can be substituted by other antimicrobial materials
such as silver or other organic compounds according to the use. The size and density of the nanomesh can be
further optimized, and the covered area can be expanded to the entire hand or other body parts. In addition,
more systematic studies on the mechanical and chemical stabilities in the long term are required to broaden
the scope of the application. Consequently, this study introduces an antimicrobial platform to combat
infectious diseases that threaten humans, particularly in the event of a pandemic, and it is expected to be
advanced toward practical use.

16
The PDMS experiments clearly demonstrated the problem with performing measurements with the sharp-tipped AFM cantilever. Young’s
modulus obtained by this technique (2.08 MPa) was 43% higher than the value acquired by the tension test. This error may be due to the
difference between the true tip radius and the typical tip radius provided by the company. The use of a flat-tipped AFM cantilever, which
compressed the PDMS beads between two parallel surfaces, effectively eliminated the error, and Young’s modulus of 1.49 MPa (similar to the
tension test result) was obtained. These results confirm that using a sharp tip for such measurements can introduce errors when evaluating
Young’s modulus, whereas using a flat tip helps avoid this problem.
To verify the reliability of the flat tip measurement method, we measured single cells of the yeast strain with the flat-tipped AFM and with a
microfluidic chip system. Both of of these processes compressed the individual yeast cells between two parallel surfaces. The small difference
in the results provided by these two methods may originate from the variation of cell properties, the error in measuring the cell diameters, or
the calibration of the AFM cantilever spring constant and on-chip force sensor. Despite these uncertainties, similar values for Young’s modulus
were obtained using the two different measurement systems described and has no obvious difference according to Student’s t-test. These
results indicate that our proposed flat-tipped method is effective for determining Young’s modulus of a whole yeast cell.
The yeast cells are also tested using the conventional AFM sharp tip method. The measurement result 0.72 MPa, which is quite different from
the flat tip measurement result in this research, as shown in Figure 6. But, it is similar with some other researches which studied yeast cell
using sharp tips (Pelling et al., 2004; Dague et al., 2010). On the other hand, the flat tip measurement result is similar with the results obtained
by some researches when the yeast cell was compressed between two flat surfaces (Arfsten et al., 2008). This result indicates that the stiffness
of the whole cell is different from the stiffness of the local point on the surface of the cell. For a PDMS bead, it is reasonable to expect a similar
result from the sharp tip and flat tip measurement. Because the bead is made of the same material and can be considered as homogeneous,
which means that the whole stiffness and the local stiffness at the surface should be the same. A yeast cell consists of cell wall, cytoplasm
filled by organelles, and so on, therefore, it is reasonable that the whole stiffness is different from the local stiffness at the surface. Thus, we
should choose proper method to study the stiffness of the target cell. If the objective is to describe the mechanical properties of a whole cell,
the flat tip should be used. While if local stiffness at the cell surface is the objective, the sharp tip should be used.
17
The PDMS experiments clearly demonstrated the problem with performing measurements with the sharp-tipped AFM cantilever. Young’s
modulus obtained by this technique (2.08 MPa) was 43% higher than the value acquired by the tension test. This error may be due to the
difference between the true tip radius and the typical tip radius provided by the company. The use of a flat-tipped AFM cantilever, which
compressed the PDMS beads between two parallel surfaces, effectively eliminated the error, and Young’s modulus of 1.49 MPa (similar to the
tension test result) was obtained. These results confirm that using a sharp tip for such measurements can introduce errors when evaluating
Young’s modulus, whereas using a flat tip helps avoid this problem.
To verify the reliability of the flat tip measurement method, we measured single cells of the yeast strain with the flat-tipped AFM and with a
microfluidic chip system. Both of of these processes compressed the individual yeast cells between two parallel surfaces. The small difference
in the results provided by these two methods may originate from the variation of cell properties, the error in measuring the cell diameters, or
the calibration of the AFM cantilever spring constant and on-chip force sensor. Despite these uncertainties, similar values for Young’s modulus
were obtained using the two different measurement systems described and has no obvious difference according to Student’s t-test. These
results indicate that our proposed flat-tipped method is effective for determining Young’s modulus of a whole yeast cell.
The yeast cells are also tested using the conventional AFM sharp tip method. The measurement result 0.72 MPa, which is quite different from
the flat tip measurement result in this research, as shown in Figure 6. But, it is similar with some other researches which studied yeast cell
using sharp tips (Pelling et al., 2004; Dague et al., 2010). On the other hand, the flat tip measurement result is similar with the results obtained
by some researches when the yeast cell was compressed between two flat surfaces (Arfsten et al., 2008). This result indicates that the stiffness
of the whole cell is different from the stiffness of the local point on the surface of the cell. For a PDMS bead, it is reasonable to expect a similar
result from the sharp tip and flat tip measurement. Because the bead is made of the same material and can be considered as homogeneous,
which means that the whole stiffness and the local stiffness at the surface should be the same. A yeast cell consists of cell wall, cytoplasm
filled by organelles, and so on, therefore, it is reasonable that the whole stiffness is different from the local stiffness at the surface. Thus, we
should choose proper method to study the stiffness of the target cell. If the objective is to describe the mechanical properties of a whole cell,
the flat tip should be used. While if local stiffness at the cell surface is the objective, the sharp tip should be used.
18
5. Discussion
5.1 Microstructure and mechanical properties of the 30Mn-1.2C-xAl steels
The novel bulk RAP approach introduced in this work provides, for the first time, a systematic evaluation of the compositional and thermo-mechanical trends
associated with a change in the Al content of a group of Triplex steels with high Mn and C concentrations. We observed that without the addition of Al to the
30Mn–1.2C steels the most favourable mechanical properties were obtained for the as-homogenised state (Figs. 2 and 3a). The observed properties are in
reasonable agreement with data reported for Mn–C alloyed TWIP steels of a similar chemical composition [27, 28]. The observed embrittlement during ageing
can be related to the formation of the coarse, pearlitic particles on the grain boundaries [27] (Fig. 5a).
High amounts of Al (~8 wt.%), on the other hand, result in pronounced strengthening during ageing, depending on the time and temperature (Figs. 2 and 3c),
and no coarse particles could be observed in this case (Fig. 5c). In the light of previous results this typical precipitation hardening behaviour, which allows
tuning of the strength and ductility, can be explained by the formation and growth of j carbides during ageing [17, 18, 21]. Due to their reportedly small size,
which is of the order of several nanometres, the j carbides could not be reliably detected or identified in the high throughput RAP OM observations conducted
in this study. The darker particles visible in Fig. 5c might be linked to j carbides.
Alloys with intermediate Al concentrations (about 2–6 wt.%) do not offer mechanical properties on the same level compared with the aforementioned
extreme cases under their respective optimal conditions (i.e. after the respective most suitable ageing treatments). On the other hand, a much smaller
influence of the ageing parameters on tensile behaviour can be observed in these cases (Figs. 2 and 3b).
Within the limitations of this study (confined range of applied heat treatments, OM investigations, etc.) this improved stability of the mechanical properties
during thermal exposure can be attributed to a concerted formation of j carbides and grain boundary pearlite, balancing the strengthening and embrittlement
effects of intermediate amounts of Al.
In general it should be underlined that detailed investigations of the role of j carbides and pearlite particles on the deformation mechanisms, as well as the
precipitation type and the structural nature of the j carbides (i.e. spinodal vs. nucleation/growth), require higher resolution techniques, such as TEM or APT.
Nonetheless, the mechanical data on both RAP (Figs. 2 and 3) and conventionally synthesised and processed alloys (Fig. 6a) are in good agreement with
previously reported values for Fe–Mn–Al–C steels [17,18]. The RAP results suggest that future efforts regarding more detailed nanostructural investigations
should focus on such high-C triplex steels with high Al concentrations (>8 wt.%), as they offer the possibility of covering the widest range of mechanical
properties via ageing treatments (scalability, Fig. 2) and exhibit the lowest possible specific weight of all such steels.

19
5. Discussion
5.1 Microstructure and mechanical properties of the 30Mn-1.2C-xAl steels
The novel bulk RAP approach introduced in this work provides, for the first time, a systematic evaluation of the compositional and thermo-mechanical trends
associated with a change in the Al content of a group of Triplex steels with high Mn and C concentrations. We observed that without the addition of Al to the
30Mn–1.2C steels the most favourable mechanical properties were obtained for the as-homogenised state (Figs. 2 and 3a). The observed properties are in
reasonable agreement with data reported for Mn–C alloyed TWIP steels of a similar chemical composition [27, 28]. The observed embrittlement during ageing
can be related to the formation of the coarse, pearlitic particles on the grain boundaries [27] (Fig. 5a).
High amounts of Al (~8 wt.%), on the other hand, result in pronounced strengthening during ageing, depending on the time and temperature (Figs. 2 and 3c),
and no coarse particles could be observed in this case (Fig. 5c). In the light of previous results this typical precipitation hardening behaviour, which allows
tuning of the strength and ductility, can be explained by the formation and growth of j carbides during ageing [17, 18, 21]. Due to their reportedly small size,
which is of the order of several nanometres, the j carbides could not be reliably detected or identified in the high throughput RAP OM observations conducted
in this study. The darker particles visible in Fig. 5c might be linked to j carbides.
Alloys with intermediate Al concentrations (about 2–6 wt.%) do not offer mechanical properties on the same level compared with the aforementioned
extreme cases under their respective optimal conditions (i.e. after the respective most suitable ageing treatments). On the other hand, a much smaller
influence of the ageing parameters on tensile behaviour can be observed in these cases (Figs. 2 and 3b).
Within the limitations of this study (confined range of applied heat treatments, OM investigations, etc.) this improved stability of the mechanical properties
during thermal exposure can be attributed to a concerted formation of j carbides and grain boundary pearlite, balancing the strengthening and embrittlement
effects of intermediate amounts of Al.
In general it should be underlined that detailed investigations of the role of j carbides and pearlite particles on the deformation mechanisms, as well as the
precipitation type and the structural nature of the j carbides (i.e. spinodal vs. nucleation/growth), require higher resolution techniques, such as TEM or APT.
Nonetheless, the mechanical data on both RAP (Figs. 2 and 3) and conventionally synthesised and processed alloys (Fig. 6a) are in good agreement with
previously reported values for Fe–Mn–Al–C steels [17,18]. The RAP results suggest that future efforts regarding more detailed nanostructural investigations
should focus on such high-C triplex steels with high Al concentrations (>8 wt.%), as they offer the possibility of covering the widest range of mechanical
properties via ageing treatments (scalability, Fig. 2) and exhibit the lowest possible specific weight of all such steels.

20
We have shown that FORCE training can take initially chaotic networks of spiking neurons and use them to mimic the natural tasks and
functions demonstrated by populations of neurons. For example, these networks were trained to learn low-dimensional dynamical systems,
such as oscillators which are at the heart of generating both rhythmic and non rhythmic motion [45]. We found FORCE training to be robust to
the spiking model employed, initial network states, and synaptic connection types.

Additionally, we showed that we could train spiking networks to display behaviors beyond low-dimensional dynamics by altering the supervisor
used to train the network. For example, we trained a statistical classifier with a network of Izhikevich neurons that could discriminate its inputs.
Extending the notion of an oscillator even further allowed us to store a complicated sequence in the form of the notes of a song, reproduce the
singing behavior of songbirds, and encode and replay a movie scene. These tasks are aided by the inclusion of a high-dimensional temporal
signal (HDTS) that discretizes time by segregating the neurons into assemblies.

FORCE training is reminiscent of how songbirds learn their stereotypical learned songs [35,46]. Juvenile songbirds are typically presented with
a species-specific song or repertoire of songs from their parents or other members of their species. These birds internalize the original
template song and subsequently use it as an error signal for their own vocalization [35–37,39,46–49]. Our model reproduced the singing
behavior of songbirds with FORCE training as the error correction mechanism. Both the spiking statistics of area RA and the song spectrogram
were accurately reproduced after FORCE training. Furthermore, we demonstrated that altering the balance between excitation and inhibition
post training degrades the singing behavior post-training. A shift to excess excitation alters the spectrogram in a highly non-linear way while a
shift to excess inhibition reduces the amplitude of all frequencies.

21
Cont’d
Inspired by the clock-like input pattern that songbirds use for learning and replay [35,36] we used a similar HDTS to encode a longer and more
complex sequence of notes in addition to a scene from a movie. We found that these signals made FORCE training faster and the subsequent
replay more accurate. Furthermore, by manipulating the HDTS frequency we found that we could speed up or reverse movie replay in a robust
fashion. We found that compressing replay resulted in higher frequency oscillations in the mean population activity. Attenuating the HDTS
decreased replay performance while transitioning the mean activity from a 4–8 Hz oscillation to a slower (≈2 Hz) oscillation. Finally, replay of
the movie was robust to lesioning neurons in the replay network.
While our episodic memory network was not associated with any particular hippocampal region, it is tempting to conjecture on how our
results might be interpreted within the context of the hippocampal literature. In particular, we found that the HDTS conferred a slow oscillation
in the mean population activity reminiscent of the slow theta oscillations observed in the hippocampus. The theta oscillation is strongly
associated to memory; however, its computational role is not fully understood, with many theories proposed [50–53]. For example, the theta
oscillation has been proposed to serve as a clock for memory formation [50,54].
Here, we show a concrete example that natural stimuli that serve as proxies for memories can be bound to an underlying oscillation in a
population of neurons. The oscillation forces the neurons to fire in discrete temporal assemblies. The oscillation (via the HDTS) can be sped up,
or even reversed resulting in an identical manipulation of the memory. Additionally, we found that reducing the HDTS input severely disrupted
replay and the underlying mean population oscillation. This mirrors experimental results that showed that theta power was predictive of
correct replay [55]. Furthermore, blocking the HDTS prevents learning and prevents accurate replay with networks trained with an HDTS
present. Blocking the hippocampal theta oscillation pharmacologically [56] or optogenetically [57] has also been found to disrupt learning.
The role of the HDTS is reminiscent of the recent discovery of time cells, which also serve to partition themselves across a time interval in
episodic memory tasks [58–60]. How time cells are formed is ongoing research however they are dependent on the medial septum, and thus
the hippocampal theta oscillation [61]. Time cells have been found in CA158, CA362 and temporally selective cells occur in the entorhinal
cortex [63].
In a broader context, FORCE trained networks could be used in the future to elucidate hippocampal functions. For example, future FORCE
trained networks can make use of biological constraints such as Dale’s law in an effort to reproduce verified spike distributions for different
neuron types with regards to the phase of the theta oscillation [64]. These networks can also be explicitly constructed to represent the
different components of the well studied hippocampal circuit. 22
Cont’d
FORCE training is a powerful tool that allows one to use any sufficiently complicated dynamical system as a basis for universal computation.
The primary difficulty in implementing the technique in spiking networks appears to be controlling the orders of magnitude between the chaos
inducing weight matrix and the feedback weight matrix. If the chaotic weight matrix is too large in magnitude (via the G parameter), the chaos
can no longer be controlled by the feedback weight matrix1. However, if the chaos inducing matrix is too weak, the chaotic system no longer
functions as a suitable reservoir. To resolve this, we derived a scaling argument for how Q should scale with G for successful training based on
network behaviors observed in ref. 1. Interestingly, the balance between these fluctuations could be related to the fading memory property, a
necessary criterion for the convergence of FORCE trained rate networks [65].
Furthermore, while we succeeded in implementing the technique in other neuron types, the Izhikevich model was the most accurate in terms
of learning arbitrary tasks or dynamics. This is due to the presence of spike frequency adaptation variables that operate on a much slower time
scale than the neuronal equations. There may be other biologically relevant forces that can increase the capacity of the network to act as a
reservoir through longer time scale dynamics, such as synaptic depression and NMDA mediated currents for example [66–68].
Furthermore, we found that the inclusion of a high-dimensional temporal signal increased the accuracy and capability of a spiking network to
reproduce long signals. In ref. 2, another type of high-dimensional supervisor is used to train initially chaotic spiking networks. Here, the
authors use a supervisor consisting of O(N2) components (see ref. 2 for more details). This is different from our approach involving the
construction of an HDTS, which serves to partition the neurons into assemblies and is of lower dimensionality than O(N2). However, from ref. 2
and our work here, increasing the dimensionality of the supervisor does aid FORCE training accuracy and capability. Finally, it is possible that
an HDTS would facilitate faster and more accurate learning in networks of rate equations and more general reservoir methods as well.
Although FORCE trained networks have dynamics that are starting to resemble those of populations of neurons, at present all top-down
procedures used to construct any functional spiking neural network need further work to become biologically plausible learning rules1,5,8. For
example, FORCE trained networks require non-local information in the form of the correlation matrix P(t). However, we should not dismiss the
final weight matrices generated by these techniques as biologically implausible simply because the techniques are themselves biologically
implausible.

23
Cont’d
Aside from the original rate formulation in ref. 1, FORCE trained rate equations have been recently applied to analyzing and reproducing
experimental data. For example, in ref. 69, the authors used a variant of FORCE training (referred to as Partial In-Network Training, PINning) to
train a rate network to reproduce a temporal sequence of activity from mouse calcium imaging data. PINning uses minimal changes from a
balanced weight matrix architecture to form neuronal sequences. In ref. 70, the authors combine experimental manipulations with FORCE
trained networks to demonstrate that preparatory activity prior to motor behavior is resistant to unilateral perturbations both experimentally,
and in their FORCE trained rate models. In ref. 71, the authors demonstrate the dynamics of reservoirs can explain the emergence of mixed
selectivity in primate dorsal Anterior Cingulate Cortex (dACC). The authors use a modified version of FORCE training to implement an
exploration/exploitation task that was also experimentally performed on primates. The authors found that the FORCE trained neurons had a
similar dynamic form of mixed selective as experimentally recorded neurons in the dACC. Finally, in ref. 72, the authors train a network of rate
neurons to encode time on the scale of seconds. This network is subsequently used to learn different spatio-temporal tasks, such as a cursive
writing task. These FORCE trained networks were able to account for psychophysical results such as Weber’s law, where the variance of a
response scales like the square of the time since the start of the response. In all cases, FORCE trained rate networks were able to account for
and predict experimental findings. Thus, FORCE trained spiking networks can prove to be invaluable for generating novel predictions using
voltage traces, spike times, and neuronal parameters.
Top-down network training techniques have different strengths and uses. For example, the Neural Engineering Framework (NEF) and spike-
based coding approaches solve for the underlying weight matrices immediately without training [5,6,8,9,11]. The solutions can be analytical as
in the spike-based coding approach, or numerical, as in the NEF approach. Furthermore, the weight matrix solutions are valid over entire
regions of the phase space, where as FORCE training uses individual trajectories as supervisors. Multiple trajectories have to be FORCE trained
into a single network to yield a comparable level of global performance over a region. Both sets of solutions yield different insights into the
structure, dynamics, and functions of spiking neural networks. For example, brain scale functional models can be constructed with NEF
networks8. Spike-based coding networks demonstrate how higher order error scaling is possible by utilizing spiking sparsely and efficiently
through balanced network solutions. While the NEF and spike-based coding approaches provide immediate weight matrix solutions, both
techniques are difficult to generalize to other types of networks or other types of tasks. Both the NEF and spike-based coding approach require
a system of closed form differential equations to determine the static weight matrix that yields the target dynamics.
In summary, we showed that FORCE can be used to train spiking neural networks to reproduce complex spatio-temporal dynamics. This
method could be used in the future to mechanically link neural activity to the complex behaviors of animals. 24
Adjusting and confirming the model for you (15 mins)
• Find an articles that provides the model research
for your study or that plays a key role in Menu for Discussion
conceptualizing or conducting your research. • State the structure of the discussion section
• Read the Discussion section of the article and • Discuss the contribution of the study
analyze the moves. Focus on the function of each • Revisit background information or literature to help
move with respect to the options in the menu of the reader
discussion section (see right). • Revisit the gap, aim, or method
• Summarize the results here: • Revisit key results

https://docs.google.com/presentation/d/1qlsGNy2 • Map onto literature (both theory and empirical

research) for comparison (support, not support)
dKnnntWu4PXuEryiPpQPfbOod_YEB6aKk-
A4/edit?usp=sharing • Identify limitations of the study and suggestions for
future research
• Restate the contribution and impact of the study
• Identify potential applications

25
Discussion vs. Conclusion
• For AAW, you may combine the two sections.
• However, please include both components even if you decide to
combine the two sections.

Discussion:
What the results mean

Conclusion:
Summary of everything that you have said in the paper
https://www.enago.com/academy/discussion-conclusion-know-difference-drafting-manuscript/
Language Tips: Modal verbs
Modal verbs:
“A modal verb is a type of verb that contextually indicates a modality such as a likelihood, ability,
permission, request, capacity, suggestion, order, obligation, or advice. Modal verbs always
accompany the base form of another verb having semantic content. ” Wikipedia (Link)

28
Modal Function Example
probability It can’t be him. James is in New York.
can ability I can write a research article.
permission/request Can I leave now?
probability It could get much hotter in August.
could past ability I could walk home in the rainy weather.
request Could I borrow your ambarella?
probability It may snow tomorrow.
may
permission You may leave now.
might probability What you’re saying might be true.
probability It should have snowed in Chicago yesterday.
should
advice You should not be writing an article.
probability His teacher must be James. I’ve seen his bicycle outside.
must obligation I must go to the court tomorrow for my jury duty.
prohibition You must not leave the class while in session.
probability The calculation has to be correct. I ran it so many times.
have to
obligation I have to finish my article by tomorrow morning. 29
 Modal Function Example
probability It can’t be him. James is in New York.
can ability I can write a research article.
permission/request Can I leave now?
probability It could get much hotter in August.
could past ability I could walk home in the rainy weather.
request Could I borrow your ambarella?
Conveying the
probability It may snow tomorrow.
right degree of may
permission You may leave now.
probability is
might probability What you’re saying might be true.
absolutely probability It should have snowed in Chicago yesterday.
essential. should
advice You should not be writing an article.
probability His teacher must be James. I’ve seen his bicycle outside.
must obligation I must go to the court tomorrow for my jury duty.
prohibition You must not leave the class while in session.
probability The calculation has to be correct. I ran it so many times.
have to
obligation I have to finish my article by tomorrow morning.
30
Conveying probability using
modal verbs
• Go to this poll: https://forms.gle/unR3i3K89mMEERDN7
• Rate the probability that each modal conveys
• We will look at the results together!

31
 Possible/Optional

This behavior may/might/can/could be specific to Homo

present sapiens.

present This behavior may/might/can/could not be specific to

(neg) Homo sapiens.

This behavior may/might/can/could have been specific to

past Homo sapiens.
This behavior may/might/can/could not have been specific
past (neg) to Homo sapiens.

32
 Expected/Likely/Probable

present This behavior should be specific to Homo sapiens.

present
This behavior should not be specific to Homo sapiens.
(neg)
past This behavior should have been specific to Homo sapiens.

This behavior should not have been specific to Homo

past (neg) sapiens.

33
 Obvious/Impossible

present This behavior must/has to be specific to Homo sapiens.

present
This behavior cannot be specific to Homo sapiens.
(neg)
past This behavior must have been specific to Homo sapiens.

This behavior cannot/could not have been specific to Homo

past (neg) sapiens.

34
Some exercise …
It is possible that the increase in temperature was due to a gas leak.
The increase in temperature ________________________________.

The parameter value will likely remain stable if we calibrate it through an iterative simulation.
The parameter value ___________________________________________________________.

It is impossible that the contamination was caused by how we measured the mineral.
The contamination _____________________________________________.

It was obvious that the bicarbonate was caused by the mineralisation of carbon.
The bicarbonate _____________________________________________________.

35
 Discussion
What to write: How to write it:
• Answer to the research question • Keep your readers in mind (i.e.,What do
they know? What do they want to know?)
• Comparison to other studies • One idea in each paragraph
• Limitations of your study • Add a concluding paragraph with
implications if you do not create a separate
• Significance/implication section for conclusion

Grammar Can your paper answer these questions?

• What is the answer to your research
• Use past tense for your findings and question?
present tense to examine the validity of
your study and suggest future directions • How do these results compare to other
published results?
• Use transitions to create coherence • What are the limitations and future
directions?
• How does your study contribute to the
field?
36
Conclusions

37
 A separate conclusion of
Results Discussion Conclusion/s
1-2 short paragraphs

The last 1-2 paragraphs of

Results Discussion
the Discussion section

A separate conclusion of
Results and Discussion Conclusion/s
1-2 short paragraphs

A long conclusion section

Results Conclusion/s
that is initially similar to Discussion
and then eventually to Conclusion

Possible scenarios
38
Let’s think! (20 mins)
• Read the following five examples of the Conclusion
section (all from UTokyo!)
• Consider each sentence or each group of sentences
and discussion its function (within the Conclusion
section)
• Post your idea here: One post per an idea (or a
function):
https://app.one.learnwiz.jp/event/6497673064

39
We developed an estimation method to estimate the average synaptic transmission
I(t) and synaptic connectivity w associated with cell assemblies from spike trains by
using mathematical models of individual neurons and cell assemblies. The
estimation ability of the proposed method was verified by simulating the neuronal
activity, and the ability to distinguish multiple cell assemblies was verified by an
electrical stimulation experiment. Multiple cell assemblies were observed by
discretely and repeatedly observing the synaptic connectivity w changing back and
forth between two patterns. Furthermore, changes were observed in approximately
twice as many cases while using the proposed method compared to when the spike
number was used, and the changes detected using the spike number were also
detected by the proposed method in every case. From these results, the proposed
method shows that it is possible to observe multiple cell assemblies based on the
biological basis of synaptic connectivity.
40
This study aimed to examine the passive cell-trapping ability of triangular microwells. The flow
circulation, which consisted of two vortices of a spanwise vortex on the top and primary streamwise
counter-rotating vortices at the leading edges of a microwell, played an important role in the selective
trapping manner. A cell crossing the front corner of the microwell would be captured. In addition, the
existence of a cell in the microwell would also change the flow pattern and help prevent the trapping
of other following cells. On the other hand, a cell crossing the side of the microwell would run across
the microwell along the trace of the primary streamwise vortices. In addition, the interaction between
the spanwise and streamwise counter-rotating vortices induces a pair of secondary counter-rotating
vortices inside the microwell. The secondary streamwise vortices would help arrange the trapping
manner, such as single, double or multiple cells, depending on the relative size of the microwells and
entrapped cells. To validate the hypothesis, polystyrene beads with a diameter of 5, 10, 15 and 20 µm
were employed with the microfluidic system at a flow rate of 0.1 mL/h with an array of 40-, 60- or
80-µm triangular microwells of a depth of 30 µm. The experimental results were in good agreement
with the theoretical ones and suggested that the size of microwells, affecting the extent of the
41
secondary streamwise vortices, and cells should be properly
We computationally optimized TP structures for single and dual absorption peaks
photodetection at arbitrary wavelengths using DE. The optimized TP structure is
composed of an Au/n-Si Schottky diode at the bottom atop a Si/SiO2 DBR
multilayer. High wavelength selectivity and near-unity absorptance of optimized TP
structures at resonance wave lengths are realized by tuning the thicknesses of each
DBR layer. Moreover, an implementable bias-controlled Schottky barrier height
method is proposed to distinguish different wavelengths for multi-peaks
photodetection. Overall, the reported optimization scheme together with the
proposed wavelength-selective detection not only offers an avenue to design high-
quality wavelength-selective Schottky barrier photodetectors, but also provides an
alternative approach to realize filter-free multi absorption peaks photodetection, and
thus open the way to novel applications related to hyperspectral imaging technology.
42
CO2 reforming of CH4 was carried out over bimetallic Co–Ni/TiO2 catalysts with different Co/Ni
ratios. The bimetallic Co–Ni/TiO2 catalysts showed highly stable activities. XRD and XPS analyses
revealed that a homogeneous alloy of cobalt and nickel was formed after the H2 reduction and
remained after the reaction. The monometallic Co/TiO2 catalyst deactivated rapidly because of the
oxidation of metal during the reaction. The small nickel substitution of cobalt (10 mol%)
dramatically improved the catalytic activity and stability. Compared with the monometallic cobalt
catalyst, the bimetallic catalysts improved the resistance to oxidation to form titanate and the
reactivity toward CH4 decomposition on the metal, giving a more reductive atmosphere over the
catalyst (e.g., H2 as a product). With the excess of nickel content (>80 mol%), the catalyst showed
higher activity for the CH4 decomposition and the reforming, but also caused more carbon formation.
With appropriate adjustment of the ratio of cobalt and nickel loading, the catalyst provides an
optimum balance between the reactions of CH4 and CO2. Long-lived Co–Ni/TiO2 catalysts that can
be used without inducing carbon formation or metal oxidation have been developed for CH4/CO2
reforming.
43
In this study, we constructed an extended arbitrary viewpoint visualization system for the teleoperation of
excavation work using a hydraulic excavator. We solved the two problems of occlusion by the arm and lack of
3D information about the excavating point. In the proposed method, it is determined whether the excavator’s
arm is included in the image, and the arbitrary viewpoint image is generated only from fish-eye camera
images that do not capture the arm. As a result, the field of view of the generated arbitrary viewpoint image
was expanded by 69.4% on average, and a visualization that contributes to safety in the teleoperation of
excavation work was realized. Furthermore, the 3D information measured by an RGB-D sensor attached to
the excavator’s arm was synthesized using an arbitrary viewpoint image. The proposed method enables
teleoperators to understand the 3D condition of the excavating point, which is necessary to determine the
position of the bucket and was not possible with previous methods. Moreover, experiments and evaluations
were conducted for each of the two solutions, and it was confirmed that an image with sufficient accuracy
could be presented in real-time. Although this study focuses on hydraulic excavators, the proposed system can
be applied to not only hydraulic excavators, but also whole machines with arms, and can provide arbitrary
viewpoint images with a wide field of view to teleoperators in real-time in the same way.
Although the suitability of the proposed method was presented by the evaluation, further verification through
actual field applications will be conducted. We plan to verify the improvement of the operability more
quantitatively by using the proposed system with actual teleoperation experiments. In addition, as described in
Section 4.3, it is important to reduce further occlusion in order to realize safer and more efficient
teleoperation, and further consideration will be required to compensate for missing images that could not be
solved by our proposed method. Furthermore, a manual or automatic control system to switch the viewpoint
of the image according to the contents of work can be considered in the development of the proposed method.
The realization of these advanced methods will lead to the development of a more practical visualization
system, which will enable safer and more efficient teleoperation of construction machines and mobile robots.
44
This paper investigated the inland flooding behavior around Faleolo International Airport, Samoa, and explored probable
drainage strategies by considering future climate change. High-spatial resolution future rainfall information (HRFR) was
estimated using satellite-based rainfall data and future projections by GCMs. Then, annual maximum daily rainfalls for
different return periods were estimated for the present and future climate. Results showed that maximum daily rainfall could
reach 245.8 mm and 432.7 mm with a return period of 125 years under the present and future climate condition,
respectively. This study applied a new method for estimation of the future rainfall. More detailed validation of the proposed
new method through comparisons with observed data is left for future tasks. Moreover, this study applied single GCM
projection. Application of multiple GCM projections is necessary to take into consideration the uncertainties of climate
change.
The estimated daily rainfall was used as forcing in a 2D overland flow model to estimate the flooding behavior around the
study area. It was found that the flooding mainly occurs to the south of the runway. The flooding area is separated by the
two taxiways into three parts. The mean value of the MID generally increases with the rainfall intensity, and parts of the
runway may be inundated under heavy rainfall.
Three drainage pipes were placed in the flooding area to discharge the rainfall water to the coast. Sensitivity analysis was
conducted to investigate the flood mitigation effect of the drainage system for different sea levels and pipe widths. The
results showed that a single drainage pipe is adequate to prevent severe inundation in catchments B and C, while multiple
drainage pipes are necessary for catchment A. Compared to the number and width of the pipeline, drainage capability was
not very sensitive to the sea level. Little difference was observed in the drainage performance if the sea level difference was
within the range of astronomical tide. Non-negligible influence, however, was observed if the water level was raised to 3 m,
which could likely occur under the scenario of sea level rise and sea level anomaly due to stormy waves. Increased width of
the pipeline showed a certain limited effect on drainage capability. Besides width, increases in the number of pipes also
showed a clear effect of reducing inundation. The location of the drainage pipes plays an important role in effectively
reducing the flooding risk. Using a larger number of drainage pipes is not always the optimum strategy for flooding
mitigation. Torrential rainfall with significantly high intensity introduces much uncertainty and raises difficulties for45the
development of drainage strategy.
 1. The gist of the paper
2. What the study has achieved
3. Background information
4. The gap or aim for the study
5. The method or approach

Menu for 6. Key results with comments

7. Implications of key results
Conclusion 8. Potential limitations
9. Potential applications
10.Potential future directions for research

46
Assignment 5 (Due: 11/15, 11:59pm)
• Finish writing the first draft of your research paper (ALL the sections).
• You can choose to have separate sections for Discussion and
Conclusion or combine them.
• No word limit but should not be too short or too long. You may want
to check a journal you may submit your article to for the guideline.
• Use the IEEE template.