Learning to distinguish cognitive subprocesses based on fMRI Tom M. Mitchell Center for Automated Learning and Discovery Carnegie Mellon University Collaborators: Luis Barrios, Rebecca Hutchinson, Marcel Just, Francisco Pereira, Jay Pujara, John Ramish, Indra Rustandi Can we distinguish brief cognitive processes using fMRI? Finds sentence ambiguous or not? Can we classify/track multiple overlapping processes? Read sentence View picture Decide whether consistent Observed fMRI: Observed button press: Mental Algebra Task [Anderson, Qin, & Sohn, 2002] 24 3 c Activity Predicted by ACT-R Model [Anderson, Qin, & Sohn, 2002] Typical ACT-R rule: IF “_ op a = b” THEN “ _ = <b <inv op> a>” [Anderson, Qin, & Sohn, 2002] Outline • Training classifiers for short cognitive processes – – – – Examples Classifier learning algorithms Feature selection Training across multiple subjects • Simultaneously classifying multiple overlapping processes – Linear Model and classification – Hidden Processes and EM Training “Virtual Sensors” of Cognitive Processes Train classifiers of form: fMRI(t, t+d) CognitiveProcess e.g., fMRI(t, t+8) = {ReadSentence, ViewPicture} • Fixed set of cognitive processes • Fixed time interval [t, t+d] Study 1: Pictures and Sentences Data from [Keller et al., 2001] View Picture Or Read Sentence t=0 4 sec. Read Sentence Or View Picture Fixation Press Button Rest 8 sec. • Subject answers whether sentence describes picture by pressing button. • 13 subjects, TR=500msec It is not true that the star is above the plus. + --* . • Learn fMRI(t,t+8) {Picture,Sentence}, for t=0,8 View Picture Or Read Sentence t=0 Fixation 4 sec. picture or sentence? Read Sentence Or View Picture Press Button Rest 8 sec. picture or sentence? Difficulties: only 8 seconds of very noisy data overlapping hemodynamic responses additional cognitive processes occuring simultaneously Learning task formulation: • Learn fMRI(t, …, t+8) {Picture, Sentence} – – – – 40 trials (40 pictures and 40 sentences) fMRI(t,…t+8) = voxels x time (~ 32,000 features) Train separate classifier for each of 13 subjects Evaluate cross-validated prediction accuracy • Learning algorithms: – – – – Gaussian Naïve Bayes Linear Support Vector Machine (SVM) k-Nearest Neighbor Artificial Neural Networks • Feature selection/abstraction – – – – Select subset of voxels (by signal, by anatomy) Select subinterval of time Summarize by averaging voxel activities over space, time … Learning a Gaussian Naïve Bayes (GNB) classifier for <f1, … fn> C For each class value, ci, 1. Estimate f1 f2 … fn 2. For each feature fj estimate modeling distribution for each ci , fj, as Gaussian, Applying GNB classifier to new instance C Support Vector Machines [Vapnik et al. 1992] • Method for learning classifiers corresponding to linear decision surface in high dimensional spaces • Chooses maximum margin decision surface • Useful in many high-dimensional domains – Text classification – Character recognition – Microarray analysis Support Vector Machines (SVM) Linear SVM Non-linear Support Vector Machines • Based on applying kernel functions to data points – Equivalent to projecting data into higher dimensional space, then finding linear decision surface – Select kernel complexity (H) to minimize ‘structural risk’ True error rate Error on training data Variance term related to kernel H complexity and number of training examples m Generative vs. Discriminative Classifiers Goal: learn , equivalently Discriminative classifier: • Learn directly Generative classifier: • Learn • Classify using Generative vs. Discriminative Classifiers Discriminative Generative What they estimate: P(C|data) P(data|C) Examples: SVM’s, Artificial Neural Nets Naïve Bayes, Bayesian networks Robustness to modeling errors Typically more robust Less robust Criterion for estimating parameters Minimize classification error Maximize data likelihood GNB vs. Logistic regression [Ng, Jordan NIPS03] Gaussian naïve Bayes Logistic regression • Model P(X|C) as a classconditional Gaussian • Model P(C|X) as a logistic function • Decision surface: hyperplane • Decision surface: hyperplane • Learning converges in O(log(n)) examples, where n is number of data attributes • Learning converges in O(n) examples • Asymptotic error less or same as GNB Accuracy of Trained Pict/Sent Classifier • Results (leave one out cross validation) – Guessing 50% accuracy – SVM: 91% mean accuracy • Single subject accuracies ranged from 75% to 98% – GNB: 84% mean accuracy – Feature selection step important for both • ~10,000 voxels x 16 time samples = 160,000 features • Selected only 240 voxels x 16 time samples Can We Train Subject-Indep Classifiers? Training Cross-Subject Classifiers for Picture/Sentence [Wang, Hutchinson, Mitchell. NIPS03] • Approach1: define “supervoxels” based on anatomically defined brain regions – Abstract to seven brain region supervoxels – Each supervoxel 100’s to 1000’s of voxels • Train on n-1 subjects, test on nth subject • Result: 75% prediction accuracy over subjects outside training set – Compared to 91% avg. single-subject accuracies – Significantly better than 50% guessing accuracy Study 2: Semantic Word Categories [Francisco Pereira] Word categories: • Fish • Trees • Vegetables • Tools • Dwellings • Building parts Experimental setup: • Block design • Two blocks per category • Each block begins by presenting category name, then 20 words • Subject indicates whether word fits category Learning task formulation • Learn fMRI(t, …, t+32) WordCategory – fMRI(t,…t+32) represented by mean fMRI image – Train on presentation 1, test on presentation 2 (and vice versa) • Learning algorithm: – 1-Nearest Neighbor, based on spatial correlation [after Haxby] • Feature selection/abstraction – Select most ‘object selective’ voxels, based on multiple regression on boxcars convolved with gamma function – 300 voxels in ventral temporal cortex produced greatest accuracy Results predicting word semantic category Mean pairwise prediction accuracy averaged over 8 subjects: • Ventral temporal: 77% (low: 57%, high 88%) • Parietal: 70% • Frontal: 67% Random guess: 50% Mean Activation per Voxel Vegetables for Word Categories P(fMRI | WordCategory) Tools one horizontal slice, ventral temporal cortex [Pereira, et al 2004] Dwellings Plot of single-voxel classification accuracies. Gaussian naïve Bayes classifier (yellow and red are most predictive). Images from three different subjects show similar regions with highly informative voxels. Subject 1 Subject 2 Subject 3 Single-voxel GNB classification error vs. p value from T-statistic N=10^6, P < 0.0001, Error = 0.51 N=10^3, P < 0.0001, Error = 0.01 Cross validated prediction error is unbiased estimate of the Bayes optimal error – the area under the intersection Question: Do different people’s brains ‘encode’ semantic categories using the same spatial patterns? No. But, there are cross-subject regularities in “distances” between categories, as measured by classifier error rates. Six-Category Study: Pairwise Classification Errors (ventral temporal cortex) * Worst * Best Subj1 Sub2 Sub3 Sub4 Sub5 Sub6 Sub7 Mean Fish Vegetables Tools Dwellings Trees Bldg Parts .20 .10 * .20 .15 .60 * .20 .15 .23 .55 * .55 * .35 * .45 * .55 .25 .55 * .46 .20 .35 .15 * .15 .25 .00 * .15 .18 .15 .20 .20 .15 .20 .30 * .25 .21 .05 * .30 .20 .05 * .15 * .05 .05 * .12 .15 .10 * .20 .25 .15 * .30 * .15 .19 LDA classification of semantic categories of photographs. [Carlson, et al., J. Cog. Neurosci, 2003] Cox & Savoy, Neuroimage 2003 Trained SVM and LDA classifiers for semantic photo categories. Classifiers applied to same subject a week later were equally accurate Lessons Learned Yes, one can train machine learning classifiers to distinguish a variety of cognitive processes – – – – Comprehend Picture vs. Sentence Read ambiguous sentence vs. unambiguous Read Noun vs. Verb Read Nouns about “tools” vs. “building parts” Failures too: – True vs. false sentences – Negative vs. affirmative sentences Which Machine Learning Method Works Best? • GNB and SVM tend to outperform KNN • Feature selection important Average per-subject classification error No Yes No Yes No Yes No Yes Which Feature Selection Works Best? Wish to learn F: <x1,x2,…xn> {A,B} • Conventional wisdom: pick features xi that best distinguish between classes A and B – E.g., sort xi by mutual information, choose the top n • Surprise: Alternative strategy worked much better The learning setting Class A Class B Voxel discriminability Voxel activity Voxel activity Rest / Fixation GNB Classifier Errors: Feature Selection feature selection method fMRI study Picture Syntactic Sentence Ambiguity Nouns vs. Verbs Word Categories .29 .26 .43 .34 .36 .36 .10 .10 Active .16 .25 .34 .08 ROI Active .18 .21 .27 .27 .31 .23 .09 NA All features Discriminate target classes ROI Active Average “Zero Signal” learning setting. Select features based on discrim(X1,X2) or discrim(Z,Xi)? Class 1 observations Class 2 observations X1=S1+N1 X2=S2+N2 Goal: learn f: XY or P(Y|X) Given: 1. Training examples <Xi, Yi> where Xi = Si + Ni , signal Si ~ P(S|Y= Yi), noise Ni ~ Pnoise Z = N0 Zero signal (fixation) 2. Observed noise with zero signal N0 ~ Pnoise “Zero Signal” learning setting Conjecture: feature selection using discrim(Z,Xi) will improve relative to discrim(X1,X2) as: • # of features increases • # of training examples decreases • signal/noise ratio decreases • fraction of relevant features decreases 2. Can we classify/track multiple overlapping processes? Input stimuli: Read sentence View picture ? Observed fMRI: Observed button press: Decide whether consistent Bayes Net related State-Space Models HMM’s, DBNs, etc. e.g., [Ghahramani, 2001] Cognitive subprocesses / state variables: fMRI: see [Hojen-Sorensen et al, NIPS99] Hidden Process Model [with Rebecca Hutchinson] Each process defined by: – ProcessID: <comprehend sentence> – Maximum HDR duration: R – EmissionDistribution: [ W(v,t) ] Interpretation Z of data: set of process instances – Desire max likelihood { <ProcessIDi, StartTimei>} – Where data likelihood is Generative model for classifying overlapping hidden processes Classifying Processes with HPMs Start time known: Start time unknown: consider candidate times S GNB classifier is a special case of HPM classifier View Picture Or Read Sentence t=0 4 sec. Read Sentence Or View Picture Fixation Press Button Rest 8 sec. GNB: picture or sentence? HPM: picture or sentence? 16 sec. picture or sentence? picture or sentence? Learning HPMs • Known start times: Least squares regression, eg. see Dale[HMB, 1999] • Unknown start times: EM algorithm – Repeat: • Estimate P(S|Y,W) • W’ arg max Currently implement M step with gradient ascent OLS learns 2 processes, overlapping in time, 1 voxel, zero noise, start times known, 10 trials [Indra Rustandi] Observed data Reconstructed data Learned process 1 Learned process 2 Estimates: -0 0.25 0.5 0.75 1 0.75 0.5 0.25 3.5108e17 -4.7535e17 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 OLS learns 2 processes, overlapping in time, 1 voxel, noise 0.2, start times known, 10 trials [Indra Rustandi] Observed data Reconstructed data Learned process 1 Learned process 2 Estimates: 0.005495 6 0.32446 0.48847 0.83317 0.99872 0.86555 0.55624 0.23633 0.050592 0.017376 0.36435 0.36134 0.4856 0.60143 0.46168 0.54137 0.47466 0.52419 Estimate Noun and Verb impulse responses [Indra Rustandi] Phase II, Words every 3 seconds. Mean LFEF, subj 08179 Verb impulse response estimated from above Verb impulse response “ground truth” from nonoverlapping stimuli Can we classify/track multiple overlapping processes? Read sentence View picture Decide whether consistent Observed fMRI: Observed button press: Learned HPM with 3 processes (S,P,D), and R=13sec (TR=500msec). Learned models S P S D? P D? S observed P S S D P P D reconstructed D start time picked to be trailStart+18 D Initial results: HPM’s on PictSent • EM chooses start time = 18 for hidden D process • Classification accuracy for heldout PS/SP trials = 15/20 = 0.75 • Heldout classification accuracy same for 2 process (P,S) and 3 process (P,S,D) models • Data likelihood over heldout data slightly better for 3 process (P,S,D) Further reading • Carlson, et al., J. Cog. Neurosci, 2003 • Cox, D.D. and R.L. Savoy, Functional magnetic resonance imaging (fMRI) ``brain reading'': detecting and classifying distributed patterns of fMRI activity in human visual cortex. NeuroImage, Volume 19, Pages 261--270, 2003. • Kjems, U., L. Hansen, J. Anderson, S. Frutiger, S. Muley, J. Sidtis, D. Rottenberg, and S. C. Strother. The quantitative evalutation of functional neuroimaging experiments: mutual information learning curves, NeuroImage 15, pp. 772--786, 2002. • Mitchell, T.M., R. Hutchinson, M. Just, S. R. Niculescu, F. Pereira, X. Wang, Classifying Instantaneous Cognitive States from fMRI Data. Proceedings of the 2003 Americal Medical Informatics Association Annual Symposium, Washington D.C., November 2003. • Mitchell, T.M., R. Hutchinson, S. R. Niculescu, F. Pereira, X. Wang, , M. Just, S. Newman. Learning to Decode Cognitive States from Brain Images, Machine Learning, 2004. • Strother S.C., J. Anderson, L.Hansen, U.Kjems, R.Kustra, J. Siditis, S. Frutiger, S. Muley, S. LaConte, and D. Rottenberg. The quantitative evaluation of functional neuroimaging experiments: The NPAIRS data analysis framework. NeuroImage 15:747-771, 2002. • Wang, X., R. Hutchinson, and T.~M. Mitchell. Training fMRI Classifiers to Detect Cognitive States across Multiple Human Subjects. Proceedings of the 2003 Conference on Neural Information Processing Systems, Vancouver, December 2003.