Demba Ba, Ph.D.


Post-doctoral Research fellow

Massachusetts Institute of Technology
Department of Brain and Cognitive Sciences

Research fellow

Harvard Medical School/Massachusetts General Hospital
Department of Anesthesia, Critical Care and Pain Medicine

Contact Information


email:                   demba(at)neurostat.mit.edu
Snail mail:           43 Vassar st, #6057

                             Cambridge, MA 02139
Office phone #:   (617) 324-1881



About

I obtained my PhD in June 2011 at MIT, in the department of Electrical Engineering and Computer Science, with a minor in Mathematics. Prof. Emery N. Brown supervised my doctoral thesis. On my doctoral thesis committee were Prof. John N. Tsitsiklis and Prof. Sanjoy Mitter. In the Fall of 2015, I will be starting as an Asssitant Professor in Electrical Engineering and Bioengineering in Harvard's School of Engineering and Applied Sciences.

Currently, I am am post-doctoral researcher in Prof. Emery N. Brown's laboratory. I am working on modelling time-varying functional connectivity in assemblies of neurons. The brain is plastic: neurons are constantly forming connections with each other. Understanding how neurons change their connectivity in response to sensory or behavioral stimuli is an important problem in neuroscience. To this end, we are developing new point-process models of time-varying networ
k connectivity which impose biophysiologically-motivated constraints on the temporal changes of the network connectivity. This research effort has led us to an exciting detour in the world of compressive sensing, where the problem of sparse recovery with temporal dynamics has not received significant attention. The inverse problems that arise in this context require the solution to large-scale optimization problem. We have developed highly-efficient, provably-convergent, iteratively re-weighted least-squares algorithms to solve structured inverse problems with temporally-smooth dynamics and sparsity in space. We have applied these algorithms to the problem of estimating structured time-frequency representations that are smooth in time and sparse in frequency, using real-valued and point-process observations. Our algorithms have led to a more precise delineation of the spectral signatures of general anesthesia induced using the drug propofol. We plan to use our framework to develop automatic systems for administering anesthetic drugs and controlling general anesthesia.


Please visit my Research and Publications pages for a summary of previous work (including my dissertation), as well as a sampling of the problems that interest me. In my CV, you will find various information pertaining to my educational experience as well as my experience in industry.

News
               06/20/14: I am excited to announce that I will be joining Harvard's School of Engineering and Applied Sciences
in                   the Fall of 2015 as an Assistant Professor in Electrical Engineering and Bioengineering. I would like to thank the                   faculty and staff at all of the great institutions that have hosted me.
               04/10/14:  Invited talk, Department of EECS, University of Michigan, Ann Arbor.

               03/20/14:  Invited talk, Department of EE, Harvard University SEAS.

               

               02/27/14:  Invited talk, Department of EE, University of Washington.

               01/23/14:  Invited talk, Department of ECE, University of California, San Diego.


               01/09/14: Our paper on theory and algorithms for simultaneous-event multivariate point-process        
               
(SEMPP) models was accepted for publication in Frontiers in Computational Neuroscience. An
               SEMPP is a multivariate point-process for which events are allowed to occur simultaneously in two
               or more of its components. These models arise in the context of assessing the importance of        
               
synchrony in neural computation. We developed a quite general framework for modelling such                                                     processes and applied it to data recorded from pairs of neurons in the rat thalamus in response to                                               periodic whisker deflections of varying velocity. We were able to demonstrate that the contribution
               of the stimulus (whisker position) to the simultaneously-occurring events can be up to an order of
               magnitude important than its contribution to non-simultaneous events, thus providing an explicit
               quantification of the importance of synchrony in the context of encoding whisker position.
         
    
              
10/05/13
: Our paper on iteratively re-weighted least-squares (IRLS) algorithms for sparse recovery
               will
appear in the IEEE Transactions on Signal Processing. The paper develops an elegant
               statistical interpretation of IRLS that simplifes global convergence analysis for a class of IRLS
               algorithms broader than previously understood in the literature. We also exhibit explicit convergence
               rates and show stability of the algorithms in the context of sparse recovery in the presence of noise!
  
 
 

               08/21/13: Our paper on likelihood methods for point-processes with refractoriness will appear in
               Neural Computation.
The paper develops a discrete-time approximation to the continuous likelihood
               of a point-process whose conditional intensity function (rate function) is piecewise Lipschitz and
               jumps to zero (refractoriness) after the occurrence of every event. We demonstrate that one can
               sample such point-process using larger bin sizes that conventional discrete-time approximations,
               while achieving the same accuracy in terms of approximating the continuous-time likelihood. In
               practice, this means that one requires less data than conventional methods to fit models with the
               same accuracy.

               03/23/13: Invited to attend MOdelling Neural Activity (MONA) workshop in Lihue, HI, June 26-28,
               2013! Presenting our work on structured time-frequency representations.


               02/06/13: Co-lecturer in MIT graduate seminar Topics in Neural Signal Processing. Course syllabus here!

               12/09/12: Invited to IPAM workshop on Adaptive Data Analysis and Sparsity.
         
              
12/02/12
: Spotlight presentation at NIPS 2012 of our work on dynamic sparse recovery.