Hardware, software and wetware: How is the brain so computationally efficient?

Speaker: Upinder S. Bhalla

Abstract

Even in the age of GPUs that run at GHz, the brain still stands out as small, fast, and efficient. To first order, a single neuron (brain cell) can be simulated by a rather substantial neural network model, or by a large system of partial differential equations. What computations can the real neuron do with these complex dynamics? I'll discuss two examples, one like the prediction of where a cricket ball will fly, and the other related to the what was that? moment when we hear a repeated sound amidst background noise. One of the uniquely powerful capabilities of neurons is their ability to extract salient signals from complex patterns in space and time, against a background of uncorrelated neural activity. The trajectory prediction problem can be framed as sequential pattern recognition. I'll discuss how sequential pattern recognition is implemented in neurons, and how this operation is equivalent to the functioning of a convolutional neural network. I'll then take on the what was that? problem of recognizing unique repeated inputs. I'll discuss some preliminary data which suggests that excitatory- inhibitory (EI) balance is asymmetric with respect to sequences of patterns. This means that certain spatiotemporal input sequences can 'escape' EI balance and cause the neuron to become more active. Overall, both these computations, take advantage of the intricate biophysics and biochemistry of the neuron to perform real-time pattern recognition with extraordinary energy efficiency.

Bio

How does the brain work? I came to this question in a roundabout way, first through an interest in physics at IIT Kanpur and Cambridge. There I realized that biology was full of amazing mechanisms, and the brain especially so. I did my PhD at Caltech, in what was then a new area of computational neuroscience. Throughout my career I have used computers both as a tool and a metaphor to understand the brain. I have been at NCBS-TIFR in Bangalore since 1996, and here I have studied how animals recognize and find odours, and how memories are formed and stored. I have been deeply involved in efforts to develop open simulation tools and neuronal data, and to promote sharing. I continue to work on the chemical signaling networks in the brain, which underlie everything from memories and computation, to growth, disease, and aging.