Wednesday 27th September 2017
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Degenerate feedback control in single neurons and networks
Neurons and neural circuits have lots of parameters that are subject to biological feedback control. Often, multiple components affect a relatively simple, global physiological variable that is then used to control the expression of each component, a scenario we call degenerate feedback. In single neurons, receptors and ion channels shape average electrical activity, and this activity itself regulates expression of receptors and channels over slow timescales. In neuronal networks, synaptic connections are shaped by reward signals during learning, while the connectivity of the network determines performance, and therefore reward. I will describe recent work that explores the consequences of degenerate feedback control in both situations. In single neurons we find that degeneracy permits flexibility in a neuron’s membrane properties at the cost of robustness to specific perturbations. In networks we see that a very minimal model of reward feedback can lead to learning in the presence of noisy synaptic turnover and local weight dynamics. This model predicts very recent experimental observations of how neural representations evolve in time, and is able to learn toy classification problems in situations where explicit error gradient methods fail.