Some Mathematical Ideas for Attacking the Brain Computer Interface Problem Michael Kirby Department of Mathematics Department of Computer Science Colorado State University Department of Mathematics Overview The Brain Computer Interface (BCI) Challenge Signal fraction analysis Takens’ theorem and classification on manifolds Nonlinear signal fraction analysis Conclusions and future work Department of Mathematics NSF BCI Group Chuck Anderson (PI), Computer Science, Colorado State Michael Kirby (Co-PI), Mathematics, Colorado State James Knight, Ph.D. Student, Colorado State Tim O’Connor, Ph.D. Student, Colorado State Ellen Curran, Medical Ethics and Jurisprudence, Dept. of Law, Keele University, Staffordshire, UK Doug Hundley, Consultant, Department of Mathematics, Whitman Pattie Davies, Occupational Therapy Department, Colorado State Bill Gavin, Dept. of Speech, Language and Hearing Sciences, University of Colorado “Geometric Pattern Analysis and Mental Task Design for a Brain-Computer Interface” Department of Mathematics SourceForge https://sourceforge.net/projects/csueeg/ Development Status: 1 - Planning Environment: Other Environment Intended Audience: Science/Research License: GNU General Public License (GPL) Natural Language: English Operating System: Linux, SunOS/Solaris Topic: Artificial Intelligence, Human Machine Interfaces, Information Analysis, Mathematics, Medical Science Apps. Department of Mathematics Chuck Anderson Department of Mathematics Pattie Davies Department of Mathematics BCI Headlines in the News Computers obey brain waves of paralyzed, Associated Press, appearing in MSNBC News, April 6, 2005 Brainwaves Control Video Games, BBC March 2004 Brainwave cap controls computer, BBC December 2004 Brain Could Guide Artificial Limbs Patients put on thinking caps, Wired News, January 2005 Monkey thoughts control computer, March 2002 Department of Mathematics Lou Gehrig’s Disease (ALS) Amyotrophic Lateral Sclerosis (ALS) , or “Locked-In Syndrome”, is an extreme neurological disorder and many patients opt against life support. Most commonly, the disease strikes people between the ages of 40 and 70, and as many as 30,000 Americans have the disease at any given time. (ALS Association website). Genetic factors appear to only account for 10 percent of all ALS cases. ALS can strike anyone, anytime. There are no effective treatments and no cure. Brain activity appears to remain vigorous while muscle control atrophies degeneritively and completely. Department of Mathematics Gulf War Veterans and ALS The following information is from a news release sent out by the Department of Veteran Affairs on December 10, 2001. (ALS Association Web posting.) “According to a news release on December 10, 2001 from the Department of Veteran Affairs, researchers conducting a large epidemiological study supported by both the Department of Veterans Affairs and the Department of Defense have found preliminary evidence that veterans who served in Desert Shield-Desert Storm are nearly twice as likely as their non-deploying counterparts to develop amyotrophic lateral sclerosis.” Department of Mathematics The Brain Computer Interface (BCI) A means for communication between person and machine via measurements associated with cerebral activity, e.g., EEG, fMRI, MEG. We assume that no muscle motion is employed such as eye twitching or finger movement. Department of Mathematics Low-Cost EEG Department of Mathematics History of EEG Duboi-Reymond (1848) reported the presence of electrical signals Caton (1875) measured “feeble” currents on the scalp Berger (1929) measured electrical signals with EEG 1930-50s EEG used in psychiatric and neurological sciences relying on visual inspection of EEG patterns 1960s-70s witness emergence of Quantitative EEG and confirmation of hemispheric specialization, e.g., left brain verbal and right brain spatial. 1980s+ observation of biofeedback Department of Mathematics Characteristics of Brainwaves Delta waves [0,4] Hz associated with sleep. Also empathy. Theta waves [4, 7.5] associated with reverie, daydreaming, meditation, creative ideas Alpha waves [7.5,12] prevalent when alert and eyes closed. Associated with relaxed positive feelings. Beta waves 12Hz+ associated with active state, eyes open. Department of Mathematics Reasons Why EEG Should Not Work for BCI Electrical activity generated by complex system of billions of neurons Brain is a “gelatinous mass” suspended in a conducting fluid Difficult to “register” electrode location Artifacts from motion, eyeblinks, swallows, heartbeat, sweating… Food, age, time of day, fatigue, motivation of subject Department of Mathematics Why EEG Can Work for BCI Many EEG studies have reported reproducible changes in brain dynamics that are task dependent! People are able to control their brainwaves via biofeedback! Department of Mathematics Biofeedback Patients may “correct” their waveforms to achieve a normal state. Kamiya demonstrated the controllability of alpha waves in 1962. Communication in morse code by turning alpha waves on and off. Stress management and sleep therapy. Move a pac-man by stimulating alpha and beta waves. Note that artifacts are a serious problem for real-time biofeedback applications. Department of Mathematics Motivation for Our Work Current biofeedback training requires 10 weeks to move a cursor. Typing requires 5 minutes/letter with 90% accuracy. Although there has been some mathematical work the field has been dominated by experiment and heuristics. Suggestions by clinical EEG experts that understanding EEG problem will have a strong mathematical component. Tremendous potential for results. Department of Mathematics EEG Data Set: Mental Tasks Resting task Imagined letter writing Mental multiplication Visualized counting Geometric object rotation Keirn and Aunon, “A new mode of communication between man and his surroundings”, IEEE Transactions on Biomedical Engineering, 37(12):1209-1214, December 1990 Department of Mathematics Lobes of the Brain Frontal Lobes Personality, emotions, problem solving. Parietal lobes Cognition, spatial relationships and mathematical abilities, nonverbal memory. Occipital lobes Vision, color, shape and movement. Temporal lobes Speech and auditory processing, language comprehension, long-term memory. Department of Mathematics Electrode Placement and Sample Data 50 C3 0 50 50 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 2000 2500 C4 0 50 50 P3 0 50 20 P4 0 20 20 O1 0 20 20 O2 0 20 Department of Mathematics 1000 1500 Sample (250 per second) Geometric Filtering of Noisy Time-Series Given a data set The Q fraction of a basis vector is defined as where Department of Mathematics Signal Fraction Optimization Determine such that D() is a maximum. Solution via the GSVD equation Department of Mathematics Department of Mathematics Original Signal SVD filter Signal fraction filter Department of Mathematics SVD basis GSVD basis 0.1 0.04 0 0.02 0. 1 0.1 0 0 0.05 0 0 0. 1 0 0 100 100 200 200 300 300 400 400 500 500 600 600 700 700 800 800 900 900 1000 1000 0.05 0 0.05 0.05 0 0 0.05 0 100 200 300 400 500 600 700 800 900 1000 0.05 0 0.05 0.2 0 0 0.05 0 100 200 300 400 500 600 700 800 900 1000 0. 2 0 0.2 0.2 0 0 0. 2 0 100 200 300 400 500 600 700 800 900 1000 0. 2 0 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 Department of Mathematics SVD reconstruction GSVD reconstruction 100 100 50 50 0 0 100 200 300 400 500 600 700 800 900 1000 0 200 200 100 100 0 0 100 200 300 400 500 600 700 800 900 1000 0 20 20 10 10 0 0 100 200 300 400 500 600 700 800 900 1000 0 30 30 20 20 10 10 0 0 100 200 300 400 500 600 700 800 900 1000 0 25 25 24.8 24.8 24.6 0 100 200 300 400 500 600 700 800 900 1000 24.6 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 Department of Mathematics Blind Signal Separation Unknown (tall) m £ n signal matrix S Unknown mixing n £ n matrix A Observed m £ n data matrix X Task: recover A and S from X alone. In general it is not possible to solve this problem. Department of Mathematics Signal Fraction Analysis Separation Theorem: The solution to the signal fraction analysis optimization problem solves the signal separation problem X = SA given 1) is observed 2) 3) In particular, Where is the solution to the GSVD problem for signal fraction analysis. Department of Mathematics Original signals (unknown) 2 5 0 0 2 0 50 100 150 200 250 300 350 400 450 500 5 5 5 0 0 5 0 50 100 150 200 250 300 350 400 450 500 5 2 5 0 0 2 5 Mixed signals (observed) 0 50 100 150 200 250 300 350 400 450 500 5 5 5 0 0 0 50 100 150 200 250 300 350 400 450 500 5 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 Department of Mathematics FastICA separation 5 2 0 0 5 0 50 100 150 200 250 300 350 400 450 500 2 2 2 0 0 2 0 50 100 150 200 250 300 350 400 450 500 2 2 5 0 0 2 5 Signal fraction separation 0 50 100 150 200 250 300 350 400 450 500 5 5 5 0 0 0 50 100 150 200 250 300 350 400 450 500 5 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 Department of Mathematics math training data letter training data 1 1 2 2 3 3 4 4 5 5 6 6 500 1000 1500 2000 2500 500 math test data 1 2 2 3 3 4 4 5 5 6 6 1000 1500 1500 2000 2500 2000 2500 letter test data 1 500 1000 2000 2500 500 Department of Mathematics 1000 1500 Artifact Removal Given the separated signals = X we may filter the ith column of by setting Where Id’ is the identity matrix with the ith row set to zero. The filtered version of the data is now Where recall the original data is Department of Mathematics Signal Fraction Filters Department of Mathematics Constructing Signal Fraction Filter Department of Mathematics Department of Mathematics Benefits of Signal Fraction Analysis Can identity sources of noise such as respirators, eyeblinks, cranial heartbeat, line noise etc… Filtering works over short periods of the signal, i.e., can remove artifacts from a time series of length 500. Can use generalizations of the signal to noise ratio to separate quantities of interest. Simple and fast to compute. Department of Mathematics Classification on Manifolds Insert slide from Istec meeting manifold: H(x) = 0 dist(A,B) large but H(A)=H(B)=0 Department of Mathematics Dynamical Systems Perspective Assume a system is described by the dynamical equations and that the solutions reside on an attracting set, e.g., a manifold. What can be said about the full system if it is only possible to observe part of the system? In the extreme, imagine we can only observe a scalar value Department of Mathematics Time Delay Embedding We may embed the scalar observable into a higher dimensional state space via the construction So now it is clear that Department of Mathematics Taken’s Theorem (simplified) Given a continuous time dynamical system with solution on a compact invariant smooth manifold M of dimension d, a continuous measurement function h(x(t)) can be time-delay embedded in to dimension 2d+1 such that there is a diffeomorphism between the embedded attractor and the actual (unobserved) solution set. Department of Mathematics The Lorenz Attractor Given a data point (x,y,z) we know which lobe by the sgn of x. But what if we only observe the z value? The lobe can be classified using Taken’s theorem and Time delay embedding. Department of Mathematics Do EEG data lie on an attractor? Department of Mathematics Elephants in the Clouds? Classification rate Random data Department of Mathematics Super Resolution Skull Caps How many electrodes are needed? 6, 16, 32, 128, 256, 512? We should be able to answer this question by means of evaluating an objective function. Through attractor reconstruction, time delay embedding techniques may practically enhance the resolution of skull caps leading to significant savings in time and equipment. Colleagues working on EEG studies in children are very enthusiastic about this! Department of Mathematics Manifolds and Nonlinear Methods (work with Fatemeh Emdad) Veronese embeddings of the data: Degree 1: (x,y) Degree 2: (x2, xy, y2) Degree 3: (x3, x2y, xy2, y3) Degree 1: (x,y,z) Degree 2: (x^2, xy, xz, y^2, yz, z^2) Degree 3: (x^3, x^2y, x^2z, xy^2, xz^2, xyz, y^3, y^2z, yz^2, z^3) Such embeddings are behind one variant of kernel SVD. Department of Mathematics Kernel SVD versus Kernel SFA Numerical Experiments: KSVD (KPCA) degree = 1, 2, 3, 4 KSFA degree = 1, 2, 3, 4 Objective: compare mode classification rates using knn for k = 1,…, 10. Department of Mathematics KSFA, KPCA degree 1 Department of Mathematics KSFA, KPCA degree 2 Department of Mathematics KSFA, KPCA degree 3 Department of Mathematics KSFA, KPCA degree 4 Department of Mathematics Relative Performance Department of Mathematics Conclusions and Future Work Present a geometric subspace approach for signal separation, artifact removal and classification. Provided evidence that brain dynamics might reside on an attractor and that time-delay embedding enhances classification rates. Illustrated a nonlinear extension to signal fraction analysis and compared with similar extension to svd. These ideas are presented in the context of EEG signals but are quite general and can be applied to images. Department of Mathematics