Computations in neuronal circuits: learning rules at "bottom-up" and "top-down connections"

25 Feb 2011

Gabriel Kreiman
Department of Opthalmology and Neurology
Children's Hospital, Boston


Seeing, hearing, feeling, planning and even proving mathematical theorems is the result of computations in the ~1011 neurons in our brains. Understanding the function of cerebral cortex is arguably one of the greatest scientific challenges of our time and will require the interplay of empirical work and rigorous theoretical models. In this talk I will give a succinct overview of theoretical approaches aimed towards understanding the biophysics of computations in cortex and focus on one particular problem that we are working on. Briefly, connections in cortex can be divided into "bottom-up" and "top-down". What type of synaptic plasticity rules at cortical top-down synapses could be compatible with the development of a stable and diverse distribution of top-down synaptic strengths? We compare two possible plasticity rules with different temporal characteristics and ask which could be a viable mechanism for learning at top-down synapses. Specifically, we compare the effects of classical spike-timing dependent plasticity (classical STDP) versus a temporally reversed version (reverse STDP). We show that a bias towards potentiation can lead to the development of strong excitatory feedback loops, which cause activity to explode but that this explosion can be prevented by a bias towards depression. Using analytical methods and simulations, we determine that only reverse STDP allows top-down connection weights to attain an unchanging distribution. Altogether, we conclude that depression-biased reverse STDP is required for the development of unchanging, diverse top-down connections.

current theory lunch schedule