Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Irina Rish, Research Staff, IBM T.J. Watson Research Center at MLconf SF - 11/13/15


Published on

Learning About Brain: Sparse Modeling and Beyond: Sparse modeling is a rapidly developing area at the intersection of statistical learning and signal processing, motivated by the age-old statistical problem of finding a relatively small subset of ”important” variables in high-dimensional datasets. Variable selection is particularly important for improving the interpretability of predictive models in scientific applications such as computational biology and neuroscience, where the main objective is to gain a better insight into functioning of a biological system, besides just learning ”black-box” predictors. Moreover, variable selection provides an effective way of avoiding the “curse of dimensionality” as it helps to prevent overfitting and reduce computational complexity in high-dimensional but relatively small-sample datasets, such as, for example, functional MRI (fMRI), where the number of variables (brain voxels) can range from 10 to 100 thousands, while the number of samples is typically limited to several hundreds.

In this talk, I will summarize our work on sparse models and other machine-learning approaches to ”brain decoding” (aka ”mind reading”), i.e. to prediction of mental states from functional MRI data, in a wide range of applications, from analyzing pain perception to discovering predictive patterns of brain activity associated with schizophrenia and cocaine addiction. I will mention several lessons learned from those applications that can hopefully generalize to other practical machine-learning problems. Finally, I will briefly discuss our recent project that focuses on inferring mental states from ”cheap” (unlike fMRI), easily collected data, such as speech and wearable sensors, with applications ranging from clinical settings (”computational psychiatry”) to everyday life (”augmented human”).

Published in: Technology
  • Be the first to comment

Irina Rish, Research Staff, IBM T.J. Watson Research Center at MLconf SF - 11/13/15

  1. 1. IBM Research © 2014 IBM Corporation © 2002 IBM Corporation IBM Research © 2015 IBM Corporation Learning about Brain: Sparse Models and Beyond Irina Rish (and many collaborators) Computational Biology Center IBM T.J. Watson Research Center, NY minx ||y - Ax||2 + λ ||x||1
  2. 2. IBM Research © 2014 IBM Corporation (An Incomplete List of) Collaborators 2 • Computational Biology Center @ IBM Watson: • Guillermo Cecchi, James Kozloski, Jeremy Rice, Laxmi Paridha, Dan He, David Haws • IBM Watson, other departments • Steve Heisig, Ravi Rao, Sasha Aravkin, Melissa Carroll (now at Google) • Neurospin (France): • JB Poline, Bertrand Thirion et al • Mt Sinai: • Rita Goldstein • Northwestern U. (Chicago) • A.V. Apkarian • SUNY Stony Brook: • Jean Honorio (now at MIT), Dimitris Samaras • Lehigh University: • Katya Scheinberg
  3. 3. IBM Research © 2014 IBM Corporation Measuring Brain Activity with Functional MRI 3 Image courtesy of fMRI Research Center at Columbia University • Blood-oxygen-level-dependent (BOLD) signal related to brain activity while subject performs some task in scanner • 4D ‘brain movie’: a sequence of 3D brain volumes 3D voxels ~ 3x3x3 mm, time repetitions (TR) ~1-2s • Challenge: high-dimensional, small-sample data 10,000 to 100,000 variables (voxels), but only 100s of TRs (samples), and less than 100 subjects
  4. 4. IBM Research © 2014 IBM Corporation Data from [Baliki, Geha, Apkarian 2008] Example: Playing Videogames Subjects play a videogame in a scanner Several types of response variables are measured: • Stimulus: Instructions, etc. • Mental states: annoyance, anxiety, happiness, etc. • ‘objective’ states: looking at faces, etc. 17 minutes For example, ‘Instructions’ variable: Can we find brain areas involved in this task? Can we predict response variables from fMRI?
  5. 5. IBM Research © 2014 IBM Corporation n Data from [Baliki, Geha, Apkarian 2008] 14 healthy subjects presented with painful thermal stimuli while in fMRI scanner, and asked to rate their pain level (using a finger-span device). Another Example: Pain Perception Where are pain-related brain areas? Can we predict pain perception and/or stimulus from fMRI?
  6. 6. IBM Research © 2014 IBM Corporation What are we looking for in fMRI data? 6 Question: given a stimulus, mental state or disorder, find relevant brain areas and/or interactions among them Traditional approach (GLM): voxel-wise correlations w/ task; too limited – mass-univariate, ignores voxel interactions! Alternative: multivariate methods to predict mental states • cognitive (e.g., viewing a picture, listening to instructions) • emotional (level of pain, anxiety, happiness) • disorders (e.g. schizophrenia, ADHD, addiction) Don’t forget to look for predictive patterns! (no ‘black-box’ predictors, please)
  7. 7. IBM Research © 2014 IBM Corporation Interpretable predictive models 7 Feature Construction: - feature engineering (e.g., network properties) [Rish et al, PLoS One 2013], [Cecchi et al, NIPS 2009] [Rish et al, SPIE Med.Imaging 2012] Feature Selection [Rish et al, SPIE Med.Imaging 2012], [Honorio et al, AISTATS 2012], [Rish et al, Brain Informatics 2010],[Carroll et al, Neuroimage 2009] - automated feature extraction (e.g., dictionary learning, deep learning, etc) [Rish et al, SfN 2011], [Rish et al, ICML 2008], ongoing work Sparse regression and sparse networks Y1 U YK … X1 XD … V1 Wk W1 ‘Biomarkers’  Predictive Features + ++ + - - - - - Predictive Model mental disorder healthy
  8. 8. IBM Research © 2014 IBM Corporation Data from [Baliki, Geha, Apkarian 2008] Mental State Prediction via Linear Regression y = Xβ + noise fMRI data (“encoding’) rows – samples (~500) Columns – voxels (~30,000) Unknown parameters (‘signal’) Measurements: mental states, behavior, tasks or stimuli Solution: embedded variable selection find a small number of (jointly) most relevant brain voxels Issue: high-dimensional, small-sample data Need to (1) prevent overfitting and (2) find interpretable solution
  9. 9. IBM Research © 2014 IBM Corporation Data from [Baliki, Geha, Apkarian 2008] LASSO and Beyond - LASSO: adds ℓ1-norm regularization (avoid overfitting) - selects relevant voxels (sparse solution  zero coefficients) Issue: no grouping of related variables! Need to augment LASSO with more ‘structure’! Group LASSO, fused LASSO, etc. Elastic Net: sparsity + grouping of correlated variables
  10. 10. IBM Research © 2014 IBM Corporation Lesson 1: Adding Proper Structure (Prior) Elastic Net vs. LASSO: • Higher prediction accuracy (close to 0.8 correlation) for pain perception, and for several video-game tasks • Better interpretability (meaningful voxel clusters vs. ‘salt and pepper’ scattered voxels) • Grouping parameter improves model stability (overlap) across different runs Some other ‘structured’ LASSO methods: • Group LASSO - when groups (e.g. regions) are known • Fused LASSO – spatial (or temporal) continuity • Moreover, adding structure in graphical LASSO, etc. [SPIE Med.Imaging 2012], [Brain Informatics 2010], [Neuroimage 2009]
  11. 11. IBM Research © 2014 IBM Corporation Learning Full-Brain Sparse Markov Networks  Problem: full-brain networks, even edge-sparse, are hard to interpret; can we identify most relevant nodes/voxels?  Proposed approach: variable (node) selection, besides the usual edge selection, using group-Lasso type of penalty Markov Networks Hypothesis: often, only a relatively few variables are interacting with each other, forming network clusters; the rest are not relevant [AISTATS 2012]
  12. 12. IBM Research © 2014 IBM Corporation Lesson 1 (cont’d): Importance of Structure (Prior)  Application: study of cocaine addicts vs. controls (Goldstein et al., 2007) performing a visual attention task with a monetary reward  Results: significantly more interpretable and statistically more accurate networks that discover most important clusters of interacting voxels cocaine subjects control subjects graphicallassoourmethod [Honorio et al, AISTATS 2012]
  13. 13. IBM Research © 2014 IBM Corporation Lesson 2: Data-Driven + Analytical Models • Dynamical model (1st order, only 3 parameters) captures inter- subject variability in pain response given stimulus • Stimulus not available? Predict from fMRI, then apply the model! VaryingPainPerception Incorporating nonlinear dynamical model into sparse learning (via hidden stimulus variable) improves over ‘direct’ sparse regression – due to very high accuracy of analytical model ! [PLoS Comp Bio 2012]
  14. 14. IBM Research © 2014 IBM Corporation Lesson 3: Go Beyond Single Best to Near-Optimal Sparse Solution Localized Activations vs ‘Holographic’ Brain Simple auditory task: localized, sparse Complex pain experience: non-sparse – ‘holographic’ Sharp transition from highly relevant first two solutions (2000 voxels), to practically irrelevant remaining voxels (0.2 and lower accuracy) No such sharp transition, slow linear decay from best (on average) 0.65 accuracy (1st solution) to 0.5 (10th sol.) and 0.4 accuracy (24th solution, 23,000 voxels removed) [SPIE Med.Imaging 2012]
  15. 15. IBM Research © 2014 IBM Corporation Schizophrenia study: discover discriminative patterns (biomarkers) Not a localized dysfunction, rather a ‘disconnection’ syndrome Explore functional network features vs local activations! No matter which classifier we used, network features outperformed local activations, thus serving as much better biomarkers. Best results: specific combination of a degree feature + classifier Network Extracted Correlation Matrix (N 2 =2x10 10 ) Thresholded Matrix MR Signal M1 V1 PP 1 N 1 N -0.5 0 0.5 1 1 N 1 N Lesson 4: Features Engineering vs Classifier Choice Voxel degrees + GMRF = 86% accuracy Top cross-voxel correlations + SVM = 93% accuracy [NIPS 2009] [PLoS ONE2013]
  16. 16. IBM Research © 2014 IBM Corporation 16 + + + + - - - Predictive Models [Cecchi et al, NIPS 2009] [Rish et al, PLOS One, 2013] [Carroll et al, Neuroimage 2009] [Scheinberg&Rish, ECML 2010] “Statistical biomarkers”: predictive patterns Schizophrenia classification: 86% to 93% accuracy [Rish et al, Brain Informatics 2010] [Rish et al, SPIE Med.Imaging 2012] [Cecchi et al, PLOS Comp Bio 2012] Cognitive state prediction in videogames: 70-95% Pain perception: 70-80% + ‘holographic’ activation discovered by sparse multivariate approach [Honorio et al, AISTATS 2012] Cocaine addiction: distinctive sparse Markov network patterns …and much more + Summary: “Mind Reading” from Functional MRI - - - -
  17. 17. IBM Research © 2014 IBM Corporation Beyond the Scanner: Using ‘Cheaper’ Sensors?
  18. 18. IBM Research © 2014 IBM Corporation Motivation: Detect Mental States, Improve Mental Function Can we avoid such tragic accidents by monitoring driver’s mental state and performing preemptive actions in real-time?
  19. 19. IBM Research © 2014 IBM Corporation Various sensors to measure physiological signals and track activities: • Heart rate • Respiration • Galvanic skin response • Accelerometer • Electroencephalogram (EEG) Yet limited inferences! Can Machine Learning bridge this gap? Wearable Devices 19 Hexoskin Heart rate Respiration Heart-rate variability Jawbone UP3 Heart rate Respiration Galvanic Skin Response (GSR) Skin temperature Ambient Temperature Accelerometer NeuroSky (EEG) Muse EEG EEG, accelerometer
  20. 20. IBM Research © 2014 IBM Corporation 20 Device detects voltage at different skin locations: 1. TP9 Behind Left Ear 2. FP1 Front Left Forehead 3. FP2 Front Right Forehead 4. TP10 Behind Right Ear FP2 FP1 TP10 TP9 Delta – Adult slow wave sleep, continuous attention processes Theta – Drowsiness, idling, inhibition Alpha – Relaxed, reflecting Beta – Alert, busy, anxious, thinking Gamma – Short term memory usage, using 2 different senses at once Muse EEG Headset
  21. 21. IBM Research © 2014 IBM Corporation Measuring Driver’s Stress Level Example: Geographic EEG Plot of relaxation index. Merging onto a highway requires extra concentration. Sensitive software would not interrupt the driver prior to and during transit of this area. EEG: Raw waveform is FFTed to power in frequency bands (e.g., from NeuroSky or Muse device) [Heisig et al, 2014]
  22. 22. IBM Research © 2014 IBM Corporation Meditation at Work (for Science’s Sake)  • Exploring differences across people engaged in the same activity • Each session consisted of 7 minutes of closed eyes meditation • Interesting clustering results  (very preliminary, though) 22 Solutions SH PB IR G HC WomenMen Green: lower than average Red: higher then average Lower beta, higher alpha: More relaxed, less anxious Higher beta, lower alpha: vice versa HC
  23. 23. IBM Research © 2014 IBM Corporation • Can you tell from EEG what kind of movie a person is watching? • 2 short (~7 min) Youtube videos, in 2 sessions • Funny (“emotional”) cat videos vs Educational (“rational”) Khan Academy • Feature extraction from raw EEG data: • 1st and 2nd order statistics from frequency bands • ‘Functional networks’ across bands and sensors • Classifiers: • SVM (RBF kernel), sparse logistic regression, decision forest, deep nets (3-layer RBM) • Best results: • Individual-level: SVM, 74.5% accuracy • Group-level: sparse logistic regression, 74% accuracy • Most-predictive features: • variances in power within frequency bands, and correlations cross-bands and cross-channels (‘functional network’) 23 Watching Movies at Work (again, for Science’s Sake) [Bashivan et al, 2015] TP9 – HG FP1 – B FP2 – A FP2 – B TP10 – G TP10 – HG
  24. 24. IBM Research © 2014 IBM Corporation Text Analytics for “Computational Psychiatry” ``Language is a window into the brain’’ - M. Covington • 93% accuracy discriminating schizophrenics from manics based on syntactic speech graphs [PLoS One, 2012] • Nearly 100% accuracy predicting 1st psychotic episode 1-2 YEARS in advance (!) via coherence and other features [Nature Schizophrenia, 2015] • 88% accuracy discriminating ecstasy and meth users from controls, using semantic features such as proximity to ‘empathy’ concept, etc., and graph features [Neuropsychopharmacology, 2014]
  25. 25. IBM Research © 2014 IBM Corporation Ongoing Work: Speech Coherence Text coherence: Currently measured as the angle between vector representations of consecutive sentences (word vectors computed by LSA) Sober vs. Non-sober Speech Coherence for Jenna Phrase-to-phrase Coherence Alternate-phraseCoherence Non-sober Sober [Heisig et al, 2014]
  26. 26. IBM Research © 2014 IBM Corporation Sensor 1 Sensor 3 I walked into a café .. Sensor 2 Sensor data  Text  Audio  Video  EEG signal  Temperature  Heart-rate  Skin- conductance Psycho- and physiological Features Voice power spectrum Text topic model Syntactic graph HRV spectrum Cheap data + Smart Analytics: Machine learning+ graph theory = Behavioral prediction Brain sciences: Psychology+ Neuroscience Behavioral Phenotype Baselining Change-point detection Predictions Towards “Augmented Human”: Real-Time Mind-Reading from Cheap Sensors
  27. 27. IBM Research © 2014 IBM Corporation References [Bashivan et al, 2015] Mental State Recognition via Wearable EEG, in Proc. of MLINI-2015 workshop at NIPS-2015. [Heisig et al, 2014] S. Heisig, G. Cecchi, R. Rao and I. Rish. Augmented Human: Human OS for Improved Mental Function. AAAI 2014 Workshop on Cognitive Computing and Augmented Human Intelligence. [Neuropsychopharmacology, 2014] A Window into the Intoxicated Mind? Speech as an Index of Psychoactive Drug Effects. Bedi G, Cecchi G A, Fernandez Slezak D, Carrillo F, Sigman M, de Wit H. Neuropsychopharmacology, 2014 [PLoS ONE, 2013] Schizophrenia as a Network Disease: Disruption of Emergent Brain Function in Patients with Auditory Hallucinations, I Rish, G Cecchi, B Thyreau, B Thirion, M Plaze, M-L Paillere-Martinot, C Martelli, J-L Martinot, J-B Poline. PloS ONE 8(1), e50625, Public Library of Science, 2013. [PLoS One, 2012] Speech Graphs Provide a Quantitative Measure of Thought Disorder in Psychosis. N.B. Mota, N.A.P. Vasconcelos, N. Lemos, A.C. Pieretti, O. Kinouchi, G.A. Cecchi, M. Copelli, S. Ribeiro. PLoS One, 2012 [SPIE Med.Imaging 2012] Sparse regression analysis of task-relevant information distribution in the brain. Irina Rish, Guillermo A Cecchi, Kyle Heuton, Marwan N Baliki, A Vania Apkarian, SPIE Medical Imaging, 2012. [AISTATS 2012] J. Honorio, D. Samaras, I. Rish, G.A. Cecchi. Variable Selection for Gaussian Graphical Models. AISTATS, 2012. [PLoS Comp Bio 2012] Predictive Dynamics of Human Pain Perception, GA Cecchi, L Huang, J Ali Hashmi, M Baliki, MV Centeno, I Rish, AV Apkarian, PLoS Comp Bio 8(10), e1002719, Public Library of Science, 2012. [Brain Informatics 2010] I. Rish, G. Cecchi, M.N. Baliki and A.V. Apkarian. Sparse Regression Models of Pain Perception, in Proc. of Brain Informatics (BI-2010), Toronto, Canada, August 2010. [NeuroImage, 2009] Prediction and interpretation of distributed neural activity with sparse models. Melissa K Carroll, Guillermo A Cecchi, Irina Rish, Rahul Garg, A Ravishankar Rao. NeuroImage 44(1), 112--122, Elsevier, 2009. [NIPS, 2009] Discriminative network models of schizophrenia, GA Cecchi, I Rish, B Thyreau, B Thirion, M Plaze, M-L Paillere- Martinot, C Martelli, J-L Martinot, J-B Poline. Advances in Neural Information Processing Systems (NIPS 2009) , pp. 252--260, 2009. 27
  28. 28. IBM Research © 2014 IBM Corporation 28 Links Publication page: Books: Practical Applications of Sparse Modeling, I Rish, GA. Cecchi, A Lozano, A Niculescu-Mizil (editors), MIT Press, 2014. I Rish and G Grabarnik. Chapman and Hall/CRC Machine Learning and Pattern Recognition, 2014.