Home
People
Publications
Talks/Lectures
Video Highlights
Links
Directions

The HAWK Seminar Series

Neural Network Control Project

Positions Available

Dynamics of Neural Networks on a Planar Patch-Clamp Array
Training, Identification, and Control

Team Members

• Marsela Jorgolli, Harvard Physics
• Alexandre Megretski, MIT LIDS
• Hongkun Park, Harvard Chemistry and Physics
• H. Sebastian Seung, MIT Brain & Cognitive Sciences, Howard Hughes Medical Institute
• Alex Shalek, Harvard
• Alec Shkolnik, MIT CSAIL
• Russ Tedrake, MIT CSAIL
• Mark Tobenkin, MIT CSAIL
• Jen Wang, MIT Brain & Cognitive Sciences
• Myung-Han Yoon, Harvard

  [Team Members Only Area]

Project Vision

Micro-fabricated patch-clamp arrays have the potential to revolutionize our understanding of learning and dynamics in networks of cultured neurons by providing simultaneous intracellular measurement and stimulation of tens to hundreds of connected neurons. These intracellular electrodes expose and control neural dynamics with temporal and spatial resolutions unparalleled by extracellular and optical methods. We are developing these patch clamp arrays atop a micro-fluidic substrate which, beyond facilitating the patch-clamp seal, can be used for spatially-targeted application and reuptake of chemical signals.

These devices enable a dream experiment. With this fantastically rich and dense instrumentation, we can probe, excite, and even reshape the dynamics of the neural network in ways that were previously impossible. Combined with a novel and rigorous concept for nonlinear system identification, we will develop high-fidelity models to describe neural activity at the network level. These models can be used to expose the topology of a cultured network, but also to investigate fundamental issues in controllability and dynamics. We intend to demonstrate our command of the network dynamics by closing a feedback loop through our instrumentation to robustly produce an activity pattern at a specified set of output neurons. Going further, we will build upon optimization theory and computational theories of reward-driven learning in the brain to operantly condition these output activity patterns. A coupling between mathematical theories of optimization and the learning dynamics of cultured neural networks would transform our understanding of the brain - providing an explanation of synaptic plasticity with close correspondence to optimization theory.

Beyond understanding biological networks, experiments with reward-driven learning can begin to reveal the brain's secrets for sequential decision making. In particular, engineering benchmarks from control theory and robotics reveal limitations in state-of-the-art optimization when the reward is nonconvex, the parameter space is large, the optimal policies are discontinuous, and/or when the dynamics of the plant contain feedback delay. With a trainable network of real neurons at our command, we have an opportunity to carefully explore nature's solution to these engineering challenges.

Publications

• Alexandre Megretski. Convex optimization in robust identification of nonlinear feedback. In Proceedings of the 47th IEEE Conference on Decision and Control, pages 1370-1374, Cancun, Mexico, Dec 9-11 2008. [ .pdf ]

• A. Starovoytov, J. Choi, and H. S. Seung. Light-directed electrical stimulation of neurons cultured on silicon wafers. J. Neurophysiology, 93:1090-1098, 2005. [ .pdf ]

• H.S. Seung. Learning in spiking neural networks by reinforcement of stochastic synaptic transmission. Neuron., 40(6):1063-73., Dec 18 2003. [ .pdf ]