Skip to main content

Michael Hughes

Followers

5

Following

4

Co-authors

2

Public Views

I am an aspiring Jurist that have fought in the United States Supreme Court and have advocated on the behalf of hundreds to be released in the notorious Chicago Cook County department of Corrections. Having great ambition for the constitution and its equity I have written over 300 pleadings and motions on the behalf of despondent detainees gaining both freedom and compensation for those injured in the purview of jurisprudence.

less

Interests

Uploads

Papers by Michael Hughes

An Open-Source Monte Carlo Ray-Tracing Simulation Tool for Luminescent Solar Concentrators with Validation Studies Employing Scattering Phosphor Films

Energies

Luminescent solar concentrators enhance the power output of solar cells through wave-guided lumin... more Luminescent solar concentrators enhance the power output of solar cells through wave-guided luminescent emission and have great potential as building-integrated photovoltaics. Luminescent solar concentrators with a variety of geometries and absorbing–emitting materials have been reported in the literature. As the breadth of available experimental configurations continues to grow, there is an increasing need for versatile Monte Carlo ray-tracing simulation tools to analyze the performance of these devices for specific applications. This paper presents the framework for a Monte Carlo ray-tracing simulation tool that can be used to analyze a host of three-dimensional geometries. It incorporates custom radiative transport models to consider the effects of scattering from luminescent media, while simultaneously modeling absorption and luminescent emission. The model is validated using experimental results for three-dimensional planar and wedge-shaped luminescent solar concentrators emplo...

Delaying U.S. Troop Exit From Afghanistan: A Policy Assessment

Academia Letters, 2021

Reliable and scalable variational inference for nonparametric mixtures, topics, and sequences

We introduce a new variational inference objective for hierarchical Dirichlet process admixture m... more We introduce a new variational inference objective for hierarchical Dirichlet process admixture models. Our approach provides novel and scalable algorithms for learning nonparametric topic models of text documents and Gaussian admixture models of image patches. Improving on the point estimates of topic probabilities used in previous work, we define full variational posteriors for all latent variables and optimize parameters via a novel surrogate likelihood bound. We show that this approach has crucial advantages for data-driven learning of the number of topics. Via merge and delete moves that remove redundant or irrelevant topics, we learn compact and interpretable models with less computation. Scaling to millions of documents is possible using stochastic or memoized variational updates.

Challenges in Computing and Optimizing Upper Bounds of Marginal Likelihood based on Chi-Square Divergences

Variational inference based on χ divergence minimization (CHIVI) provides a way to approximate a ... more Variational inference based on χ divergence minimization (CHIVI) provides a way to approximate a model’s posterior while obtaining an upper bound on the marginal likelihood. However, in practice CHIVI relies on Monte Carlo (MC) estimates of an upper bound objective that at modest sample sizes are not guaranteed to be true bounds on the marginal likelihood. This paper provides an empirical study of CHIVI performance on a series of synthetic inference tasks. We show that CHIVI is far more sensitive to initialization than classic VI based on KL minimization, often needs a very large number of samples (over a million), and may not be a reliable upper bound. We also suggest possible ways to detect and alleviate some of these pathologies, including diagnostic bounds and initialization strategies.

MuscleDB : An open source cloud based platform for visualizing RNA-seq data

Book Division of Pulmonary and Critical Care Medicine

Sampling From a 1-D Truncated Normal Distribution : Theory , Algorithms , and Applications

The 1-D truncated Normal distribution is a one-dimensional probability density function proportio... more The 1-D truncated Normal distribution is a one-dimensional probability density function proportional to the 1-D Normal (Gaussian) distribution, except with support truncated to some interval a < x < b rather than the entire real line. Sampling from this distribution is straight-forward, except when the truncated interval occurs far from the mean. In this report, we discuss efficient and exact sampling techniques for the one-sided truncation x > a even as a → ∞. We develop a sampling algorithm appropriate for all possible truncation points a, and discuss its practical utility in variety of applications such as censored data tasks, probit regression, and multinomial regression.

National park user pays systems in Australia. Cost recovery vs access for all

This paper discusses the tension between user pays systems as a means of cost recovery and equity... more This paper discusses the tension between user pays systems as a means of cost recovery and equity of public access using the example of Australian national parks. Six Australian state based national park management agencies were interviewed about their user pays systems. Australian national parks are managed as a public good, for biodiversity conservation, and to provide for recreation and tourism opportunities. Legislated social equity requirements result in discounted user pays systems that are generally not cost effective. This raises the question of whether user pays systems for cost recovery are compatible with access management based on social equity.

Rethinking clinical prediction: Why machine learning must consider year of care and feature aggregation

ArXiv, 2018

Machine learning for healthcare often trains models on de-identified datasets with randomly-shift... more Machine learning for healthcare often trains models on de-identified datasets with randomly-shifted calendar dates, ignoring the fact that data were generated under hospital operation practices that change over time. These changing practices induce definitive changes in observed data which confound evaluations which do not account for dates and limit the generalisability of date-agnostic models. In this work, we establish the magnitude of this problem on MIMIC, a public hospital dataset, and showcase a simple solution. We augment MIMIC with the year in which care was provided and show that a model trained using standard feature representations will significantly degrade in quality over time. We find a deterioration of 0.3 AUC when evaluating mortality prediction on data from 10 years later. We find a similar deterioration of 0.15 AUC for length-of-stay. In contrast, we demonstrate that clinically-oriented aggregates of raw features significantly mitigate future deterioration. Our su...

Supervised Topic Models for Video Activity Recognition

Topic models successfully capture latent structure useful for unsupervised analysis of bag-of-wor... more Topic models successfully capture latent structure useful for unsupervised analysis of bag-of-words data. Applying these models to domains such as video activity recognition requires two critical extensions: (1) incorporating supervised information (activity labels) to recover topic structure with greater discriminative power and (2) moving beyond the bag-of-words assumption to model temporal dynamics. We propose two parallel investigations to accomplish these tasks. First, we will study generic supervision techniques for topic models, exposing shortcomings in previously published generative approaches and exploring new discriminative models based on Mixtures of Experts. Second, we will apply these supervised models to video activity classification on the challenging Hollywood2 and Olympic Sports datasets, and explore extensions that capture chronological structure inherent in real-world activities.

Rapid Model Comparison by Amortizing Across Models

Comparing the inferences of diverse candidate models is an essential part of model checking and e... more Comparing the inferences of diverse candidate models is an essential part of model checking and escaping local optima. To enable efficient comparison, we introduce an amortized variational inference framework that can perform fast and reliable posterior estimation across models of the same architecture. Our Any Parameter Encoder (APE) extends the encoder neural network common in amortized inference to take both a data feature vector and a model parameter vector as input. APE thus reduces posterior inference across unseen data and models to a single forward pass. In experiments comparing candidate topic models for synthetic data and product reviews, our Any Parameter Encoder yields comparable posteriors to more expensive methods in far less time, especially when the encoder architecture is designed in model-aware fashion.

Ecocean: conservation through technological innovation

Variational Inference for Beta-Bernoulli Dirichlet Process Mixture Models

A commonly used paradigm in diverse application areas is to assume that an observed set of indivi... more A commonly used paradigm in diverse application areas is to assume that an observed set of individual binary features is generated from a Bernoulli distribution with probabilities varying according to a Beta distribution. In this paper, we present our nonparametric variational inference algorithm for the Beta-Bernoulli observation model. Our primary focus is clustering discrete binary data using the Dirichlet process (DP) mixture model.

Scalable Adaptation of State Complexity for Nonparametric Hidden Markov Models

Bayesian nonparametric hidden Markov models are typically learned via fixed truncations of the in... more Bayesian nonparametric hidden Markov models are typically learned via fixed truncations of the infinite state space or local Monte Carlo proposals that make small changes to the state space. We develop an inference algorithm for the sticky hierarchical Dirichlet process hidden Markov model that scales to big datasets by processing a few sequences at a time yet allows rapid adaptation of the state space cardinality. Unlike previous point-estimate methods, our novel variational bound penalizes redundant or irrelevant states and thus enables optimization of the state space. Our birth proposals use observed data statistics to create useful new states that escape local optima. Merge and delete proposals remove ineffective states to yield simpler models with more affordable future computations. Experiments on speaker diarization, motion capture, and epigenetic chromatin datasets discover models that are more compact, more interpretable, and better aligned to ground truth segmentations tha...

Refinery: An Open Source Topic Modeling Web Platform

J. Mach. Learn. Res., 2017

We introduce Refinery, an open source platform for exploring large text document collections with... more We introduce Refinery, an open source platform for exploring large text document collections with topic models. Refinery is a standalone web application driven by a graphical interface, so it is usable by those without machine learning or programming expertise. Users can interactively organize articles by topic and also refine this organization with phrase-level analysis. Under the hood, we train Bayesian nonparametric topic models that can adapt model complexity to the provided data with scalable learning algorithms. The project website http://daeilkim.github.io/refinery/ contains Python code and further documentation.

bnpy : Reliable and scalable variational inference for Bayesian nonparametric models

We introduce bnpy, a new inference engine implemented in Python for unsupervised learning from mi... more We introduce bnpy, a new inference engine implemented in Python for unsupervised learning from millions of examples. Our framework applies to a large class of parametric and Bayesian nonparametric (BNP) clustering models that capture sequential, hierarchical, or spatial structure. For BNP models, we develop memoized variational algorithms that explore adding or removing clusters to discover compact, interpretable models. Python code: http://bitbucket.org/michaelchughes/bnpy-dev/

Surf Zone Turbulence and Suspended Sediment Dynamics—A Review

Journal of Marine Science and Engineering, 2021

The existence of sandy beaches relies on the onshore transport of sand by waves during post-storm... more The existence of sandy beaches relies on the onshore transport of sand by waves during post-storm conditions. Most operational sediment transport models employ wave-averaged terms, and/or the instantaneous cross-shore velocity signal, but the models often fail in predictions of the onshore-directed transport rates. An important reason is that they rarely consider the phase relationships between wave orbital velocity and the suspended sediment concentration. This relationship depends on the intra-wave structure of the bed shear stress and hence on the timing and magnitude of turbulence production in the water column. This paper provides an up-to-date review of recent experimental advances on intra-wave turbulence characteristics, sediment mobilization, and suspended sediment transport in laboratory and natural surf zones. Experimental results generally show that peaks in the suspended sediment concentration are shifted forward on the wave phase with increasing turbulence levels and i...

Taming fNIRS-based BCI Input for Better Calibration and Broader Use

The 34th Annual ACM Symposium on User Interface Software and Technology, 2021

Brain-computer interfaces (BCI) are an emerging technology with many potential applications. Func... more Brain-computer interfaces (BCI) are an emerging technology with many potential applications. Functional near-infrared spectroscopy (fNIRS) can provide a convenient and unobtrusive real time input for BCI. fNIRS is especially promising as a signal that could be used to automatically classify a user's current cognitive workload. However, the data needed to train such a classifier is currently not widely available, difficult to collect, and difficult to interpret due to noise and cross-subject variation. A further challenge is the need for significant user-specific calibration. To address these issues, we introduce a new dataset gathered from 15 subjects and a new multi-stage supervised machine learning pipeline. Our approach learns from both observed data and augmented data derived from multiple subjects in its early stages, and then fine-tunes predictions to an individual subject in its last stage. We show promising gains in accuracy in a standard "n-back" cognitive workload classification task compared to baselines that use only subject-specific data or only group-level data, even when our approach is given much less subject-specific data. Even though these experiments analyzed the data retrospectively, we carefully removed anything from our process that could not have been done in real time, because our process is targeted at future real-time operation. This paper contributes a new dataset, a new multi-stage training pipeline, results showing significant improvement compared to alternative pipelines, and discussion of the implications for user interface design. Our complete dataset and software are publicly available at https://tuftshci-lab.github.io/code_and_datasets/. We hope these results make * Both authors contributed equally to this research. † Both authors jointly supervised this work. This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike International 4.0 License.

Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, 2017

Expressive classifiers such as neural networks are among the most accurate supervised learning me... more Expressive classifiers such as neural networks are among the most accurate supervised learning methods in use today, but their opaque decision boundaries make them difficult to trust in critical applications. We propose a method to explain the predictions of any differentiable model via the gradient of the class label with respect to the input (which provides a normal to the decision boundary). Not only is this approach orders of magnitude faster at identifying input dimensions of high sensitivity than sample-based perturbation methods (e.g. LIME), but it also lends itself to efficiently discovering multiple qualitatively different decision boundaries as well as decision boundaries that are consistent with expert annotation. On multiple datasets, we show our approach generalizes much better when test conditions differ from those in training.

CT-guided aspiration of suspected pancreatic infection

International journal of pancreatology, 1995

We have performed CT-guided percutaneous needle aspiration in 104 patients with severe pancreatit... more We have performed CT-guided percutaneous needle aspiration in 104 patients with severe pancreatitis strongly suspected of harboring pancreatic infection on the basis of systemic toxicity and CT findings (Balthazar CT grade D or E). Of these 104 patients, 51 (49%) were documented with pancreatic infection. Gram stain was positive in 54 of 58 infected aspirates, and culture was positive in all 58. Klebsiella, Escherichia coli, and Staphylococcus aureus were the most frequent organisms. Eighty-six percent of infected processes contained only one organism. Overall, pancreatic infection was documented by GPA within the first 2 wk in approx one-half of patients. There were no complications. The overall rate of infection decreased from 60 (1980-1987) to 34% (1988-1995) (p = 0.011). This change was caused by a reduction in the rate of infected necrosis from 67 to 32% (p = 0.015). The overall mortality rate remained at 20%. The mortality of sterile pancreatitis was not different from infected pancreatitis (p = 0.14). We conclude that GPA is a safe, accurate method of diagnosis of pancreatic infection. The rate of pancreatic infection appears to be decreasing. The overall mortality of severe pancreatitis among patients suspected of harboring pancreatic infection has remained unchanged because of the high mortality associated with both infected necrosis and severe sterile necrosis.

Urinary trypsinogen activation peptide (TAP) predicts severity in patients with acute pancreatitis

International journal of pancreatology, 1997

Conclusions. Urinary TAP obtained within the first 48 h of the onset of symptoms can distinguish ... more Conclusions. Urinary TAP obtained within the first 48 h of the onset of symptoms can distinguish patients with severe acute pancreatitis. Background. Urinary trypsinogen activation peptide (TAP) has recently been described as an early marker of severity in acute pancreatitis. Methods. In a multicenter study, urine samples were collected for TAP concentration at 6-12, 24, and 48 h after admission from 139 patients with acute pancreatitis (99 with mild disease, 40 with severe disease) and from 50 control patients. Severity of acute pancreatitis was defined by the presence of organ failure and/ or pancreatic necrosis on dynamic contrast-enhanced computed tomography. Results. Median urinary TAP in the 139 patients with acute pancreatitis compared to the 50 control patients was significantly higher at admission, 4.6 vs 0.8 ng/mL (p < 0.001), and 6-12 h, 1.9 vs 0.55 ng/mL (p = 0.04). Among patients who presented within 48 h of the onset of symptoms, the median urinary TAP for severe pancreatitis (9 patients) compared to mild pancreatitis (40 patients) was significantly higher at admission, 29.6 vs. 3.6 ng/mL (p = 0.001). Also, when obtained within 48 h of the onset of symptoms, all patients with severe pancreatitis had an admission urinary TAP level > 10 ng/mL. The sensitivity and specificity of an admission urinary TAP >10 tbr severe pancreatitis was 100 and 85%, respectively. Given a cutoffof 10 ng/mL for an admission urinary TAP obtained within 48 h of the onset of symptoms, the negative predictive value was 100% for mild pancreatitis.

An Open-Source Monte Carlo Ray-Tracing Simulation Tool for Luminescent Solar Concentrators with Validation Studies Employing Scattering Phosphor Films

Energies

Luminescent solar concentrators enhance the power output of solar cells through wave-guided lumin... more Luminescent solar concentrators enhance the power output of solar cells through wave-guided luminescent emission and have great potential as building-integrated photovoltaics. Luminescent solar concentrators with a variety of geometries and absorbing–emitting materials have been reported in the literature. As the breadth of available experimental configurations continues to grow, there is an increasing need for versatile Monte Carlo ray-tracing simulation tools to analyze the performance of these devices for specific applications. This paper presents the framework for a Monte Carlo ray-tracing simulation tool that can be used to analyze a host of three-dimensional geometries. It incorporates custom radiative transport models to consider the effects of scattering from luminescent media, while simultaneously modeling absorption and luminescent emission. The model is validated using experimental results for three-dimensional planar and wedge-shaped luminescent solar concentrators emplo...

Delaying U.S. Troop Exit From Afghanistan: A Policy Assessment

Academia Letters, 2021

Reliable and scalable variational inference for nonparametric mixtures, topics, and sequences

We introduce a new variational inference objective for hierarchical Dirichlet process admixture m... more We introduce a new variational inference objective for hierarchical Dirichlet process admixture models. Our approach provides novel and scalable algorithms for learning nonparametric topic models of text documents and Gaussian admixture models of image patches. Improving on the point estimates of topic probabilities used in previous work, we define full variational posteriors for all latent variables and optimize parameters via a novel surrogate likelihood bound. We show that this approach has crucial advantages for data-driven learning of the number of topics. Via merge and delete moves that remove redundant or irrelevant topics, we learn compact and interpretable models with less computation. Scaling to millions of documents is possible using stochastic or memoized variational updates.

Challenges in Computing and Optimizing Upper Bounds of Marginal Likelihood based on Chi-Square Divergences

Variational inference based on χ divergence minimization (CHIVI) provides a way to approximate a ... more Variational inference based on χ divergence minimization (CHIVI) provides a way to approximate a model’s posterior while obtaining an upper bound on the marginal likelihood. However, in practice CHIVI relies on Monte Carlo (MC) estimates of an upper bound objective that at modest sample sizes are not guaranteed to be true bounds on the marginal likelihood. This paper provides an empirical study of CHIVI performance on a series of synthetic inference tasks. We show that CHIVI is far more sensitive to initialization than classic VI based on KL minimization, often needs a very large number of samples (over a million), and may not be a reliable upper bound. We also suggest possible ways to detect and alleviate some of these pathologies, including diagnostic bounds and initialization strategies.

MuscleDB : An open source cloud based platform for visualizing RNA-seq data

Book Division of Pulmonary and Critical Care Medicine

Sampling From a 1-D Truncated Normal Distribution : Theory , Algorithms , and Applications

The 1-D truncated Normal distribution is a one-dimensional probability density function proportio... more The 1-D truncated Normal distribution is a one-dimensional probability density function proportional to the 1-D Normal (Gaussian) distribution, except with support truncated to some interval a < x < b rather than the entire real line. Sampling from this distribution is straight-forward, except when the truncated interval occurs far from the mean. In this report, we discuss efficient and exact sampling techniques for the one-sided truncation x > a even as a → ∞. We develop a sampling algorithm appropriate for all possible truncation points a, and discuss its practical utility in variety of applications such as censored data tasks, probit regression, and multinomial regression.

National park user pays systems in Australia. Cost recovery vs access for all

This paper discusses the tension between user pays systems as a means of cost recovery and equity... more This paper discusses the tension between user pays systems as a means of cost recovery and equity of public access using the example of Australian national parks. Six Australian state based national park management agencies were interviewed about their user pays systems. Australian national parks are managed as a public good, for biodiversity conservation, and to provide for recreation and tourism opportunities. Legislated social equity requirements result in discounted user pays systems that are generally not cost effective. This raises the question of whether user pays systems for cost recovery are compatible with access management based on social equity.

Rethinking clinical prediction: Why machine learning must consider year of care and feature aggregation

ArXiv, 2018

Machine learning for healthcare often trains models on de-identified datasets with randomly-shift... more Machine learning for healthcare often trains models on de-identified datasets with randomly-shifted calendar dates, ignoring the fact that data were generated under hospital operation practices that change over time. These changing practices induce definitive changes in observed data which confound evaluations which do not account for dates and limit the generalisability of date-agnostic models. In this work, we establish the magnitude of this problem on MIMIC, a public hospital dataset, and showcase a simple solution. We augment MIMIC with the year in which care was provided and show that a model trained using standard feature representations will significantly degrade in quality over time. We find a deterioration of 0.3 AUC when evaluating mortality prediction on data from 10 years later. We find a similar deterioration of 0.15 AUC for length-of-stay. In contrast, we demonstrate that clinically-oriented aggregates of raw features significantly mitigate future deterioration. Our su...

Supervised Topic Models for Video Activity Recognition

Topic models successfully capture latent structure useful for unsupervised analysis of bag-of-wor... more Topic models successfully capture latent structure useful for unsupervised analysis of bag-of-words data. Applying these models to domains such as video activity recognition requires two critical extensions: (1) incorporating supervised information (activity labels) to recover topic structure with greater discriminative power and (2) moving beyond the bag-of-words assumption to model temporal dynamics. We propose two parallel investigations to accomplish these tasks. First, we will study generic supervision techniques for topic models, exposing shortcomings in previously published generative approaches and exploring new discriminative models based on Mixtures of Experts. Second, we will apply these supervised models to video activity classification on the challenging Hollywood2 and Olympic Sports datasets, and explore extensions that capture chronological structure inherent in real-world activities.

Rapid Model Comparison by Amortizing Across Models

Comparing the inferences of diverse candidate models is an essential part of model checking and e... more Comparing the inferences of diverse candidate models is an essential part of model checking and escaping local optima. To enable efficient comparison, we introduce an amortized variational inference framework that can perform fast and reliable posterior estimation across models of the same architecture. Our Any Parameter Encoder (APE) extends the encoder neural network common in amortized inference to take both a data feature vector and a model parameter vector as input. APE thus reduces posterior inference across unseen data and models to a single forward pass. In experiments comparing candidate topic models for synthetic data and product reviews, our Any Parameter Encoder yields comparable posteriors to more expensive methods in far less time, especially when the encoder architecture is designed in model-aware fashion.

Ecocean: conservation through technological innovation

Variational Inference for Beta-Bernoulli Dirichlet Process Mixture Models

A commonly used paradigm in diverse application areas is to assume that an observed set of indivi... more A commonly used paradigm in diverse application areas is to assume that an observed set of individual binary features is generated from a Bernoulli distribution with probabilities varying according to a Beta distribution. In this paper, we present our nonparametric variational inference algorithm for the Beta-Bernoulli observation model. Our primary focus is clustering discrete binary data using the Dirichlet process (DP) mixture model.

Scalable Adaptation of State Complexity for Nonparametric Hidden Markov Models

Bayesian nonparametric hidden Markov models are typically learned via fixed truncations of the in... more Bayesian nonparametric hidden Markov models are typically learned via fixed truncations of the infinite state space or local Monte Carlo proposals that make small changes to the state space. We develop an inference algorithm for the sticky hierarchical Dirichlet process hidden Markov model that scales to big datasets by processing a few sequences at a time yet allows rapid adaptation of the state space cardinality. Unlike previous point-estimate methods, our novel variational bound penalizes redundant or irrelevant states and thus enables optimization of the state space. Our birth proposals use observed data statistics to create useful new states that escape local optima. Merge and delete proposals remove ineffective states to yield simpler models with more affordable future computations. Experiments on speaker diarization, motion capture, and epigenetic chromatin datasets discover models that are more compact, more interpretable, and better aligned to ground truth segmentations tha...

Refinery: An Open Source Topic Modeling Web Platform

J. Mach. Learn. Res., 2017

We introduce Refinery, an open source platform for exploring large text document collections with... more We introduce Refinery, an open source platform for exploring large text document collections with topic models. Refinery is a standalone web application driven by a graphical interface, so it is usable by those without machine learning or programming expertise. Users can interactively organize articles by topic and also refine this organization with phrase-level analysis. Under the hood, we train Bayesian nonparametric topic models that can adapt model complexity to the provided data with scalable learning algorithms. The project website http://daeilkim.github.io/refinery/ contains Python code and further documentation.

bnpy : Reliable and scalable variational inference for Bayesian nonparametric models

We introduce bnpy, a new inference engine implemented in Python for unsupervised learning from mi... more We introduce bnpy, a new inference engine implemented in Python for unsupervised learning from millions of examples. Our framework applies to a large class of parametric and Bayesian nonparametric (BNP) clustering models that capture sequential, hierarchical, or spatial structure. For BNP models, we develop memoized variational algorithms that explore adding or removing clusters to discover compact, interpretable models. Python code: http://bitbucket.org/michaelchughes/bnpy-dev/

Surf Zone Turbulence and Suspended Sediment Dynamics—A Review

Journal of Marine Science and Engineering, 2021

The existence of sandy beaches relies on the onshore transport of sand by waves during post-storm... more The existence of sandy beaches relies on the onshore transport of sand by waves during post-storm conditions. Most operational sediment transport models employ wave-averaged terms, and/or the instantaneous cross-shore velocity signal, but the models often fail in predictions of the onshore-directed transport rates. An important reason is that they rarely consider the phase relationships between wave orbital velocity and the suspended sediment concentration. This relationship depends on the intra-wave structure of the bed shear stress and hence on the timing and magnitude of turbulence production in the water column. This paper provides an up-to-date review of recent experimental advances on intra-wave turbulence characteristics, sediment mobilization, and suspended sediment transport in laboratory and natural surf zones. Experimental results generally show that peaks in the suspended sediment concentration are shifted forward on the wave phase with increasing turbulence levels and i...

Taming fNIRS-based BCI Input for Better Calibration and Broader Use

The 34th Annual ACM Symposium on User Interface Software and Technology, 2021

Brain-computer interfaces (BCI) are an emerging technology with many potential applications. Func... more Brain-computer interfaces (BCI) are an emerging technology with many potential applications. Functional near-infrared spectroscopy (fNIRS) can provide a convenient and unobtrusive real time input for BCI. fNIRS is especially promising as a signal that could be used to automatically classify a user's current cognitive workload. However, the data needed to train such a classifier is currently not widely available, difficult to collect, and difficult to interpret due to noise and cross-subject variation. A further challenge is the need for significant user-specific calibration. To address these issues, we introduce a new dataset gathered from 15 subjects and a new multi-stage supervised machine learning pipeline. Our approach learns from both observed data and augmented data derived from multiple subjects in its early stages, and then fine-tunes predictions to an individual subject in its last stage. We show promising gains in accuracy in a standard "n-back" cognitive workload classification task compared to baselines that use only subject-specific data or only group-level data, even when our approach is given much less subject-specific data. Even though these experiments analyzed the data retrospectively, we carefully removed anything from our process that could not have been done in real time, because our process is targeted at future real-time operation. This paper contributes a new dataset, a new multi-stage training pipeline, results showing significant improvement compared to alternative pipelines, and discussion of the implications for user interface design. Our complete dataset and software are publicly available at https://tuftshci-lab.github.io/code_and_datasets/. We hope these results make * Both authors contributed equally to this research. † Both authors jointly supervised this work. This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike International 4.0 License.

Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, 2017

Expressive classifiers such as neural networks are among the most accurate supervised learning me... more Expressive classifiers such as neural networks are among the most accurate supervised learning methods in use today, but their opaque decision boundaries make them difficult to trust in critical applications. We propose a method to explain the predictions of any differentiable model via the gradient of the class label with respect to the input (which provides a normal to the decision boundary). Not only is this approach orders of magnitude faster at identifying input dimensions of high sensitivity than sample-based perturbation methods (e.g. LIME), but it also lends itself to efficiently discovering multiple qualitatively different decision boundaries as well as decision boundaries that are consistent with expert annotation. On multiple datasets, we show our approach generalizes much better when test conditions differ from those in training.

CT-guided aspiration of suspected pancreatic infection

International journal of pancreatology, 1995

We have performed CT-guided percutaneous needle aspiration in 104 patients with severe pancreatit... more We have performed CT-guided percutaneous needle aspiration in 104 patients with severe pancreatitis strongly suspected of harboring pancreatic infection on the basis of systemic toxicity and CT findings (Balthazar CT grade D or E). Of these 104 patients, 51 (49%) were documented with pancreatic infection. Gram stain was positive in 54 of 58 infected aspirates, and culture was positive in all 58. Klebsiella, Escherichia coli, and Staphylococcus aureus were the most frequent organisms. Eighty-six percent of infected processes contained only one organism. Overall, pancreatic infection was documented by GPA within the first 2 wk in approx one-half of patients. There were no complications. The overall rate of infection decreased from 60 (1980-1987) to 34% (1988-1995) (p = 0.011). This change was caused by a reduction in the rate of infected necrosis from 67 to 32% (p = 0.015). The overall mortality rate remained at 20%. The mortality of sterile pancreatitis was not different from infected pancreatitis (p = 0.14). We conclude that GPA is a safe, accurate method of diagnosis of pancreatic infection. The rate of pancreatic infection appears to be decreasing. The overall mortality of severe pancreatitis among patients suspected of harboring pancreatic infection has remained unchanged because of the high mortality associated with both infected necrosis and severe sterile necrosis.

Urinary trypsinogen activation peptide (TAP) predicts severity in patients with acute pancreatitis

International journal of pancreatology, 1997

Conclusions. Urinary TAP obtained within the first 48 h of the onset of symptoms can distinguish ... more Conclusions. Urinary TAP obtained within the first 48 h of the onset of symptoms can distinguish patients with severe acute pancreatitis. Background. Urinary trypsinogen activation peptide (TAP) has recently been described as an early marker of severity in acute pancreatitis. Methods. In a multicenter study, urine samples were collected for TAP concentration at 6-12, 24, and 48 h after admission from 139 patients with acute pancreatitis (99 with mild disease, 40 with severe disease) and from 50 control patients. Severity of acute pancreatitis was defined by the presence of organ failure and/ or pancreatic necrosis on dynamic contrast-enhanced computed tomography. Results. Median urinary TAP in the 139 patients with acute pancreatitis compared to the 50 control patients was significantly higher at admission, 4.6 vs 0.8 ng/mL (p < 0.001), and 6-12 h, 1.9 vs 0.55 ng/mL (p = 0.04). Among patients who presented within 48 h of the onset of symptoms, the median urinary TAP for severe pancreatitis (9 patients) compared to mild pancreatitis (40 patients) was significantly higher at admission, 29.6 vs. 3.6 ng/mL (p = 0.001). Also, when obtained within 48 h of the onset of symptoms, all patients with severe pancreatitis had an admission urinary TAP level > 10 ng/mL. The sensitivity and specificity of an admission urinary TAP >10 tbr severe pancreatitis was 100 and 85%, respectively. Given a cutoffof 10 ng/mL for an admission urinary TAP obtained within 48 h of the onset of symptoms, the negative predictive value was 100% for mild pancreatitis.