The Seung Lab

Justin Werfel

Graduate Student
(617) 252-1739
jkwerfel@mit.edu

Mechanisms of Motor Learning in Birdsong Production

Listening to other birds, songbirds memorize a temporally extended pattern of sound. Later they learn to reproduce that song, beginning with babbling and gradually learning to manipulate their vocal apparatus to generate the desired result. Neural recordings have been made from various areas of the brain believed to be involved with aspects of song including: motor control, at levels from individual acoustic features up to entire syllables; auditory feedback; comparison of that feedback with the memorized acoustic pattern; and so on. How are the neurons within each area connected to produce the observed activity, and just what computation does that pattern of activity accomplished? Under what learning rule are those connections established? And how are the different areas connected together, so that the high-level behavior and learning we observe emerge? I am particularly interested in models of the overall motor learning process, unifying what is known about the different motor areas and their activity patterns under different conditions. Existing unified models tend to contradict observations or lack important detail, but in light of currently emerging results from experimental labs, a more satisfying model may be possible.

Reinforcement Learning in Neural Networks

Many low-level models for learning are based on neural networks as their substrate. Standard frameworks for classification, regression, or trajectory learning for such networks, such as the Perceptron learning rule and backpropagation, suffer from problems of biological realism. It can be difficult to take a rule that may be simple to write down, and realize it in a network constructed from neurons as we understand them; or the standard prescription may require the network to run both forwards and backwards in time, an ability hard to imagine in living cells. I have been studying the class of REINFORCE algorithms introduced by Ronald Williams, which can demonstrate performance comparable to that of standard algorithms like backpropagation, but lack many of the same objections to the plausibility of their implemention in a network of independent biological or artificial processing units.

EEG for Telepathy and Mind Control

No, not really. This project (which does involve EEG) is in its inception; actual details will be forthcoming at a later time.