In this one-day workshop we will explore perspectives, novel methodology, and its impact at the interface of Machine Learning, Inference, Neuroimaging and Neuroscience. We aim to bring researchers from the machine learning and neuroscience communities together, in order to discuss open questions, identify the core points for a number of the controversial issues, and discuss the interaction among novel methodology and neuroscientific problems. The special topic of this year is "Going Beyond the Scanner", which includes making inference about the subject's mental states from ''cheap'' data such as subject's speech and/or text, audio, video, EEG and other wearable devices.

The workshop will be structured around these main topics:

I) machine learning and pattern recognition methodology in brain research
II) functional connectivity and dynamical models of brain activity
III) multi-modal analysis including mental state inference from behavioral data
IV) linking machine learning, neuroimaging and neuroscience
V) beyond the scanner methods to complement neuroimaging

Each session will be include invited talks or original , and an in depth discussion. This will be followed by original contributions. Original contributions will also be presented and discussed during a poster session. Each workshop day will end with a panel discussion, during which we will address specific questions, and invited speakers will open each segment with a brief presentation of their opinion.

Location: Room 515 a

Friday, December 11, 2015

Morning Session:

 8:45-9:00  opening remarks (organizers) 
 9:00-10:00  Invited talk: Gabriel Kreiman (Harvard) - Visual pattern completion: from neural circuits to computational models
 10:00-10:30  Contributed talk: Brian Helfer - Assessing Functional Neural Connectivity as an Indicator of Cognitive Performance
 10:30-11:00  coffee break, poster setup (poster session 1)
 11:00-11:30  poster spotlights  
 12:00-12:30  poster session 

12:30-2:30 lunch

Afternoon Session:

 2:30-3:30 Invited talk: Mitsuo Kawato (ATR) - Spectrum of Psychiatric Disorders revealed by Machine Learning Algorithms
 Contributed talk:  Romy Lorenz - Stopping criteria for boosting automatic experimental design using real-time fMRI with Bayesian optimization
 4:00-4:30 coffee break and continuation of poster session
 4:30-5:30 Invited talk: Sylvain Baillet (McGill) - Possible mechanisms enabling functional brain connectivity in the resting and active states
 5:30-6:30 panel discussion

Saturday, December 12, 2015

Morning Session:

 9:00-10:00  Invited talk: John Anderson (CMU) - The Sequential Structure of Thought.
 10:00-10:30  Contributed talk: Mattia Rigotti - Estimating the dimensionality of neural responses with fMRI Repetition Suppression
 10:30-11:00  coffee break, poster setup (poster session 2)
 11:00-11:30  Contributed talk: Dirk Walther - A Bayesian Test for Comparing Classifier Errors
 12:00-12:30  poster spotlights

12:30-2:30 lunch

Afternoon Session:

 2:30-3:30  Invited talk:  Cheryl Corcoran (Columbia)
 contributed talk: Sami Remes - Classification of weak multi-view signals by sharing factors in a mixture of Bayesian group factor analyzers
 4:00-4:30 coffee break/poster session 2
 4:30-5:30 Invited talk: Tom Mitchell (CMU)
 5:30-6:30 panel discussion + closing remarks

Invited Talks:

Gabriel Kreiman: Visual pattern completion: from neural circuits to computational models

Pattern completion is a fundamental cornerstone of cognition and intelligence. Extrapolating from partial information is essential in multiple domains including sensory processing (e.g. identifying partially occluded objects), social inference (e.g. capturing the gist of interactions from minimal information), or language processing (e.g. interpreting narratives). I will describe our initial steps towards understanding the mechanisms underlying pattern completion in the context of visual recognition. Evidence from invasive neural recordings along the human ventral visual stream, behavioral reaction time and masking experiments and computational simulations suggest that top-down and/or recurrent processing play a fundamental role in orchestrating pattern completion.

Mitsuo Kawato: Spectrum of Psychiatric Disorders revealed by Machine Learning Algorithms

Joint work with Jun Morimoto, Masahiro Yamashita, Hiroshi Imamizu

Current diagnosis and therapy of psychiatric disorders with childhood and adult onsets are categorical, as described by DSM-5. However, a large GWAS study identified genetic risk loci shared by several disorders, and a meta-analysis found decreases common to several disorders in gray matter volumes of specific brain regions; therefore, more emphasis has recently been placed on the spectrum of disorders and an exploration of biological dimensions to characterize it. We formed a consortium (SRPBS-DecNef project, supported by Japanese AMED) in 2013 to develop diagnosis and therapy based on computational neuroscience and machine learning algorithms. We developed biomarkers of autism spectrum disorder (ASD), schizophrenia, major depressive disorder (MDD), and obsessive-compulsive disorder (OCD) utilizing resting-state functional connectivity MRI (rs-fcMRI). For this work, about 1,000 samples were collected from more than 10 scanners in Japan. We aimed to develop new therapies using decoded fMRI real-time neurofeedback (DecNef) based on multi-voxel pattern analysis (Shibata et al., Science 2011) or connectivity real-time neurofeedback (FCNef) based on rs-fcMRI biomarkers (Megumi et al., Front Human Neurosci 2015). Here, we introduce a few studies supporting a spectrum of psychiatric disorders.

   Yahata et al. (under revision) successfully developed an ASD classifier with 82% correct for three Japanese sites and a marked generalization capability with 75% correct for the USA ABIDE dataset. A machine-learning algorithm consisted of a nested feature selection by L1-SCCA and cross validation by sparse logistic regression, and it selected only 16 functional connections (FCs) from 9,730 FCs among 140 brain regions. The same algorithm was also applied to schizophrenia, MDD and OCD with comparable performances. The four classifiers altogether selected about 60 FCs, and the 1,000 samples were hierarchically clustered within these 60-FC coordinates with 5 clusters. The hierarchical structure turned out to be ((((OCD, Scz), ASD), Healthy), MDD), and each cluster corresponded to diagnostic labels with 75–89%; the overall distribution of samples formed a spectrum with significant overlaps between disorders.

   Yamashita et al. (2015 Sci Rep) predicted a learning plateau of a verbal 3-back working memory task in 17 healthy young Japanese participants by a linear weighted summation of 16 FCs among 18 functional brain networks. This Working Memory Prediction Model based on Healthy Young Japanese participants (WMPMHYJ) explained individual working memory capabilities of the much larger and more greatly diversified USA dataset. WMPMHYJ also predicted individual differences in working memory ability in schizophrenic patients of the SRPBS-DecNef consortium. It is known that the severity of working memory deficit is in the order of schizophrenia, MDD, OCD and ASD. WMPMHYJ reproduced not only this order but also the quantitative aspects of deficits in different psychiatric disorders. If the neural mechanisms connecting functional brain networks with cognitive (dys-)functions in different psychiatric disorders were different, any model that was developed using only healthy participants, such as WMPMHYJ, would not have generalized to multiple disorders. Therefore, the spectrum view of psychiatric disorders is supported at the level that connects macro-functional connectome with cognitive functions. There seems to be a common neural mechanism connecting functional networks with syndromes, and only FCs might be characteristically different among various psychiatric disorders.

Sylvain Baillet - Possible mechanisms enabling functional brain connectivity in the resting and active states

John Anderson:
The Sequential Structure of Thought.

There has been great progress in mining the spatial structure of brain activation to  understand cognition. However, complex cognition plays out over time and has a rich temporal structure as well. I will describe our efforts to combine temporal and spatial pattern matching in fMRI, EEG, and MEG. We have been able to combine bottom-up discovery processes with top-down guidance from a rich cognitive model.

Contributed Talks:

Title:Classification of weak multi-view signals by sharing factors in a mixture of Bayesian group factor analyzers
Abstract:We propose a novel classification model for weak signal data, building upon a recent model for Bayesian multi-view learning, Group Factor Analysis (GFA). 
Instead of assuming all data to come from a single GFA model, we allow latent clusters, each having a different GFA model and producing a different class distribution. We show that sharing information across the clusters, by sharing factors, increases the classification accuracy considerably; the shared factors essentially form a flexible noise model that explains away the part of data not related to classification. Motivation for the setting comes from single-trial functional brain imaging data, having a very low signal-to-noise ratio and a natural multi-view setting, with the different sensors, measurement modalities (EEG, MEG, fMRI) and possible auxiliary information as views. We demonstrate our model on a MEG dataset.
Authors:Sami Remes*, Aalto University
Tommi Mononen, Aalto University
Samuel Kaski, Aalto University

Title:A Bayesian Test for Comparing Classifier Errors
Abstract:Multi-class classification algorithms have become an important tool for the analysis of neuroimaging data. Classification errors contain potentially important information that often goes unreported. It is therefore desirable to quantitatively compare patterns of errors between different experimental conditions. Here we present a Bayesian test that is based on comparing evidence in favor of two competing hypotheses, one stating dependence and one stating independence of two given error patterns. We derive analytical solutions for the likelihoods of both hypotheses. We compare the results from our new test with two other methods of comparing error patterns using data from an fMRI experiment and we substantiate reasons for adopting our proposal and for future work.
Authors:Dirk Walther*, University of Toronto
Emanuele Olivetti, Bruno Kessler Foundation

Title:Estimating the dimensionality of neural responses with fMRI Repetition Suppression
Abstract:We propose a novel method that exploits fMRI Repetition Suppression (RS-fMRI) to measure the dimensionality of the set response vectors, i.e. the dimension of the space of linear combinations of neural population activity patterns in response to specific task conditions. RS-fMRI measures the overlap between response vectors even in brain areas displaying no discernible average differential BOLD signal. We show how this property can be used to estimate the neural response dimensionality in areas lacking macroscopic spatial patterning. The importance of dimensionality derives from how it relates to the a neural circuit's functionality. As we will show, the dimensionality of the response vectors is predicted to be high in areas involved in multi-stream integration, while it is low in areas where the inputs from independent sources do not interact or merely overlap linearly. Our method can be used to identify and functionally characterize cortical circuits that integrate multiple independent information pathways.
Authors:Mattia Rigotti, IBM T.J. Watson Research Center
Stefano Fusi*, Columbia University

Title:Stopping criteria for boosting automatic experimental design using real-time fMRI with Bayesian optimization
Abstract:Bayesian optimization has recently been proposed as a practical and efficient tool through which to tune parameters in many difficult settings. Recently, such techniques have been combined with real-time fMRI to propose a novel framework which turns on its head the conventional functional neuroimaging approach. This closed-loop method automatically designs the optimal experiment to evoke a desired target brain pattern. One of the challenges associated with extending such methods to real-time brain imaging is the need for adequate stopping criteria, an aspect of Bayesian optimization which has received limited attention. In light of high scanning costs and limited attentional capacities of subjects an accurate and reliable stopping criteria is essential. In order to address this issue we propose and empirically study the performance of two stopping criteria.
Authors:Romy Lorenz*, Imperial College London
Ricardo Monti, Imperial College London
Ines Violante, Imperial College London
christoforos Anagnostopoulos, Imperial College London
Aldo Faisal, Imperial College London
Giovanni Montana, King's College London
Robert Leech, Imperial College London

Title:Assessing Functional Neural Connectivity as an Indicator of Cognitive Performance
Abstract:Studies in recent years have demonstrated that neural organization and structure impact an individual’s ability to perform a given task. Specifically, individuals with greater neural efficiency have been shown to outperform those with less organized functional structure. In this work, we compare the predictive ability of properties of neural connectivity on a working memory task. We provide two novel approaches for characterizing functional network connectivity from electroencephalography (EEG), and compare these features to the average power across frequency bands in EEG channels. Our first novel approach represents functional connectivity structure through the distribution of eigenvalues making up channel coherence matrices in multiple frequency bands. Our second approach creates a connectivity network at each frequency band, and assesses variability in average path lengths and degree across the network. Failures in digit and sentence recall on single trials are detected using a Gaussian classifier for each feature set, at each frequency band. The classifier results are then fused across frequency bands, with the resulting detection performance summarized using the area under the receiver operating characteristic curve (AUC) statistic. Fused AUC results of 0.63/0.58/0.61 for digit recall failure and 0.58/0.59/0.54 for sentence recall failure are obtained from the connectivity structure, graph variability, and channel power features respectively.
Authors:Brian Helfer*, MIT Lincoln Laboratory
James Williamson, MIT Lincoln Laboratory
Benjamin Miller, MIT Lincoln Laboratory
Joseph Perricone, MIT Lincoln Laboratory
Thomas Quatieri, MIT Lincoln Laboratory

Posters - Day 1:

Title:Improving our understanding of transfer-learning in ERP based BCI
Abstract:Neuroimaging is challenging for machine learning methods as data is high dimensional, noisy, contaminated by artefacts and non-stationary. Substantial inter-subject variability complicates matters e.g.~for a decoder in a brain-computer interface (BCI) system. Tackling it requires either subject-specific decoding models, or inter-subject transfer learning. Considering the large inter-subject discrepancies observed, the success of transfer learning for BCIs seems counter-intuitive. We show that inter-subject variability in the sensor space does not necessarily imply that a common latent structure across subjects cannot exist. We argue that transfer learning might be beneficial even though visual inspection of the data would lead one to believe otherwise. This will be illustrated with toy data and verified using relevant dimensionality estimation (RDE) on data of an auditory ERP-BCI experiment.
Authors:Pieter-Jan Kindermans*, TU-Berlin
Michael Tangermann, 
Martijn Schreuder, 
Mikio Braun, 
Klaus-Robert Müller,

Title:Locally-Optimized Inter-Subject Alignment of Functional Cortical Regions
Abstract:Inter-subject registration of cortical areas is necessary in functional imaging (fMRI) studies for making inferences about equivalent brain function across a population. However, many high-level visual brain areas are defined as peaks of functional contrasts whose cortical position is highly variable. As such, most alignment methods fail to accurately map functional regions of interest (ROIs) across participants. To address this problem, we propose a locally optimized registration method that directly predicts the location of a seed ROI on a separate target cortical sheet by maximizing the functional correlation between their time courses, while simultaneously allowing for non-smooth local deformations in region topology. Our method outperforms the two most commonly used alternatives (anatomical landmark-based AFNI alignment and cortical convexity-based FreeSurfer alignment) in overlap between predicted region and functionally-defined LOC. Furthermore, the maps obtained using our method are more consistent across subjects than both baseline measures. Critically, our method represents an important step forward towards predicting brain regions without explicit localizer scans and understanding the poorly understood relationship between the location of functional regions, their anatomical extent, and the consistency of computations those regions perform across people.
Authors:Marius Cătălin Iordan*, Stanford University
Armand Joulin, Stanford University
Diane Beck, University of Illinois at Urbana-Champaign
Li Fei-Fei, Stanford University

Title:Semantic vector space models predict neural responses to complex visual stimuli
Abstract:Encoding models have as their objective to predict neural responses to naturalistic stimuli with the aim of elucidating how sensory information is represented in the brain. This prediction is achieved by representing the stimulus in terms of a suitable feature space and using this feature space to linearly predict observed neural responses. Here, we investigate to what extent semantic vector space models can be used to predict neural responses to complex visual stimuli. We show that these models provide good predictions of neural responses in downstream visual areas, improving significantly over a low-level control model based on Gabor wavelet pyramids. The outlined approach provides a new way to model and map high-level semantic representations across cortex.
Authors:Umut Güçlü*, Radboud University
Marcel van Gerven, Radboud University

Title:Nonparametric Modeling of Dynamic Functional Connectivity in fMRI Data
Abstract:Dynamic functional connectivity (FC) has in recent years become a topic of interest in the neuroimaging community. Several models and methods exist for both functional magnetic resonance imaging (fMRI) and electroencephalography (EEG), and the results point towards the conclusion that FC exhibits dynamic changes. The existing approaches modeling dynamic connectivity have primarily been based on time-windowing the data and k-means clustering. We propose a non-parametric generative model for dynamic FC in fMRI that does not rely on specifying window lengths and number of dynamic states. Rooted in Bayesian statistical modeling we use the predictive likelihood to investigate if the model can discriminate between a motor task and rest both within and across subjects. We further investigate what drives dynamic states using the model on the entire data collated across subjects and task/rest. We find that the number of states extracted are driven by subject variability and preprocessing differences while the individual states are almost purely defined by either task or rest. This questions how we in general interpret dynamic FC and points to the need for more research on what drives dynamic FC.
Authors:Søren Nielsen*, Technical University Denmark
Kristoffer Madsen, Danish Research Centre for Magnetic Resonance
Rasmus Røge, Technical University of Denmark
Mikkel Schmidt, Technical University of Denmark
Morten Mørup, Technical University of Denmark

Title:Resting state brain networks from EEG: Comparing hidden Markov states with classical microstates
Abstract:Functional brain networks exhibit dynamics on the sub-second temporal scale and are often assumed to embody the physiological substrate of cognitive processes. 
Here we analyse the temporal and spatial dynamics of these states, as measured by EEG, with a hidden Markov model and compare this approach to classical EEG microstate analysis. We find dominating state lifetimes of 100-150 ms for both approaches. The state topographies show obvious similarities. However, they also feature distinct spatial and especially temporal properties. 
These differences may carry physiological meaningful information originating from patterns in the data that the HMM is able to integrate while the microstate analysis is not. This hypothesis is supported by a consistently high pairwise correlation of the temporal evolution of EEG microstates which is not observed for the HMM states and which seems unlikely to be a good description of the underlying physiology. 
However, further investigation is required to determine the robustness and the functional and clinical relevance of EEG HMM states in comparison to EEG microstates.
Authors:Tammo Rukat*, University of Oxford
Adam Baker, Oxford Centre for Human Brain Activity
Andrew Quinn, Oxford Centre for Human Brain Activity
Mark Woolrich, Oxford Centre for Human Brain Activity

Title:Neural Spatial Consistency of hierarchical vision models
Abstract:The human visual system is assumed to use multiple stages of 
computation to process visual information. Similarly, hierarchical stages of computation are employed by models such as HMAX, Bag of Words (BoW) and Convolutional Neural Networks(CNN). These models are increasingly being compared using the human brain. However, hierarchical models come in a variety of different forms and vary in essential characteristics such as the number of stages in the model and the computations that are applied within each stage. While these models have been tested in isolation, they have not been tested against each other in terms of their neural spatial architecture. In this paper we use Neural Spatial Consistency Analysis (NSCA) to compare how the different computational steps in these models explain brain responses. We used BOLD fMRI data from 20 subjects who watched a 11 minute natural movie. Voxel-wise we employed a distance based-variation partitioning against the dissimilarity matrices of the models to determine the neural spatial architecture. The correlation between the neural spatial architecture gives us to what extent the different computational stages of the models correspond to each other. We find that the different stages of CNN explains brain responses from low-level brain areas to higher regions of the brain. Both HMAX and BoW explain brain responses primarily in the early visual cortex. Additionally both HMAX and BoW are highly correlated to layer 3 of CNN, further providing evidence that that both these models only capture low level information. Comparing BoW and HMAX with CNN we note that BoW has a stronger correlation to CNN than HMAX. This leads us to believe that though HMAX and CNN are biologically inspired, BoW is more similar to CNN than HMAX.
Authors:Kandan Ramakrishnan*, University of Amsterdam
H.Steven Scholte, University of Amsterdam
Arnold M.Smeulders, University of Amsterdam
Sennay Ghebreab, University of Amsterdam

Posters - Day 2:

Title:Convolutional Deep Neural Network for Stereotypical Motor Movement Recognition in Autism
Abstract:Autism Spectrum Disorders (ASD) are often associated with specific atypical postural or motor behaviors, of which Stereotypical Motor Movements (SMMs) have a specific visibility. While the identification and the quantification of SMM patterns remain complex, its automation would provide support to accurate tuning of the intervention in the therapy of autism. Therefore, it is essential to develop automatic SMM detection systems in a real world setting, taking care of strong inter-subject and intra-subject variability. Wireless accelerometer sensing technology can provide a valid infrastructure for real-time SMM detection, however such variability remains a problem also for machine learning methods, in particular whenever handcrafted features extracted from accelerometer signal are considered. Here, we propose to employ the deep learning paradigm in order to learn discriminating features from multi-sensor accelerometer signals. Our results provide preliminary evidence that feature learning and transfer learning embedded in the deep architecture achieve higher accurate SMM detectors in longitudinal scenarios.
Authors:Nastaran Mohammadian Rad*, University of Trento
Seyed Mostafa Kia, University of Trento
Andrea Bizzego, 
Giuseppe Jurman, 
Paola Venuti, 
Cesare Furlanello,

Title:Predicting Intentions in Uncontrolled Environment Using 4-channel Dry Electrode EEG
Abstract:Current EEG systems have used dense gel-electrode systems for BCI applications. These EEG systems lack mobility and robustness, thus making brain-computer interface applications in daily-life scenarios inconvenient or unfeasible. Recent advancements in terms of miniaturization and wearability have made it possible for measurements to be performed outside the laboratory environment. In this paper we investigate classification of motor intentions using a wearable EEG system in a regular office environment. Accuracies on par with current state-of-the-art have been achieved using IMEC's wearable EEG system with 4 active-dry channels with commonly used feature extraction and classification techniques. Our results demonstrate that with such a setup, movement intentions can be predicted 460 ms before the movement onset with accuracy greater than 80%.
Authors:Ulf Grossekathoefer*, Holst Centre\IMEC
Paruthi Pradhapan, Holst Centre\IMEC
Vojkan Mihajlovic, Holst Centre\IMEC

Title:A hierarchical Bayesian framework for modeling individual differences in mental processing stages with a hidden semi-Markov model
Abstract:Conventionally, electroencephalogram (EEG) trials are averaged either time-locked to the stimulus or the response to obtain event-related potentials (ERP), which are components that capture the time occurrence of important cognitive events in a task. Built upon the previous work of a hidden semi-Markov model (HSMM) that dynamically captures the trial-to–trial variability of cognitive events during EEG recordings, the sequence of mental processing stages that one goes through in an association recognition task can be recovered. It remains to be investigated, however, if the durations of any of the mental processing stages differ from individual to individual during the experiment, and if there are multiple stages that co-vary across individuals. As a result, individual difference is modeled to better understand the temporal dynamics of mental processing stages. A hierarchical Bayesian framework is a tradeoff between pooling all individuals together and modeling them separately, in the sense that there are different parameterizations at the individual level. At the same time, a global constraint is implemented where the individual parameters are sampled from a population-level distribution with hyperparameters. Previously widely used in modeling individual differences over behavior data, hierarchical Bayesian framework can now be applied to EEG data through the integration with the stage durations of the HSMM. The HSMM and the hierarchical Bayesian parameters can then be learned simultaneously via Gibbs sampling. It is observed that there is very little individual difference for the duration of visual encoding stage; and that individuals who spend longer time in the memory retrieval stage demonstrate a faster decision-making stage.
Authors:Qiong Zhang*, Carnegie Mellon University
John Anderson , Carnegie Mellon University 
Robert Kass, Carnegie Mellon University

Title:Emotional Intensity analysis in Bipolar subjects
Abstract:The massive availability of digital repositories of human thought opens radical novel way of studying the human mind. 
Natural language processing tools and computational models have evolved such that many mental conditions are predicted by analysing speech. 
Transcription of interviews and discourses are analyzed using syntactic, grammatical or sentiment analysis to infer the mental state. 
Here we set to investigate if classification of Bipolar and control subjects is possible. 
We develop the Emotion Intensity Index based on the Dictionary of Affect, and find that subjects categories are distinguishable. 
Using classical classification techniques we get more than 75\% of labeling performance. 
These results sumed to previous studies show that current automated speech analysis is capable of identifying altered mental states towards a quantitative psychiatry. 
Authors:Facundo Carrillo, 
Natalia Mota, 
Mauro Copelli, 
Sidarta Ribeiro, 
Mariano Sigman, 
Guillermo Cecchi, 
Diego Fernandez Slezak*, Universidad de Buenos Aires

Title:Mental State Recognition via Wearable EEG
Abstract:The increasing quality and affordability of consumer electroencephalogram (EEG) headsets make them attractive for situations where medical grade devices are impractical. Predicting and tracking cognitive states is possible for tasks that were previously not conducive to EEG monitoring. For instance, monitoring operators for states inappropriate to the task (e.g. drowsy drivers), tracking mental health (e.g. anxiety) and productivity (e.g. tiredness) are among possible applications for the technology. Consumer grade EEG headsets are affordable and relatively easy to use, but they lack the resolution and quality of signal that can be achieved using medical grade EEG devices. Thus, the key questions remain: to what extent are wearable EEG devices capable of mental state recognition, and what kind of mental states can be accurately recognized with these devices? In this work, we examined responses to two different types of input: instructional (‘logical’) versus recreational (‘emotional‘) videos, using a range of machine-learning methods. We tried SVMs, sparse logistic regression, and Deep Belief Networks, to discriminate between the states of mind induced by different types of video input, that can be roughly labeled as ‘logical’ vs. ‘emotional’. Our results (around 75% accuracy) demonstrate a significant potential of wearable EEG devices in differentiating cognitive states between situations with large contextual but subtle apparent differences.
Authors:Pouya Bashivan*, University of memphis
Irina Rish, IBM T. J. Watson Research Center
Steve Heisig, IBM T. J. Watson Research Center