Signal recovery in cortico-cortical communication through representational learning in the brain

Abstract:

The talk will discuss different objectives of representational learning in the brain by describing the results from two studies. The first study investigates if the objective of efficient coding can explain the shape diversity of receptive fields found in V1 (Ringach, 2002). Early models of efficient sparse coding (e.g. Olshausen & Field 1996) were unable to reproduce the full diversity of shapes observed in V1. We developed network models for V1 that differed in two respects from the early models. First, the learned representations are overcomplete, that is, the cortical neurons in the model outnumber the dimension of the thalamic input. Second, the sparseness constraint forces the number of active neurons to be small, in contrast to limiting the average neuronal activity as done in the earlier models. The new model accurately predicted the distribution of shapes of cortical receptive fields found in nature suggesting that efficient coding is crucial, however, the high dimensionality and the type of sparseness are other critical features of the model (Rehn & Sommer, 2007).

The second study takes off where the first ends. How can high-dimensional sparse representations formed in V1 be transmitted through long-range fiber projections to other brain regions? The problem is that the number of axons sent out from one region to another is much smaller than the number of local neurons in each region (Schuez et al. Cerebral Cortex, 2006). Combining ideas from sparse coding and compressive sampling, we have discovered a learning algorithm, adaptive compressive sampling, that can be proven to establish and maintain lossless communication through fiber bottlenecks. Our discovery can explain how a neural population in the brain targeted by an axonal fiber projection can use the arriving signals to learn response properties that not only convey the full information sent into the projection but also resemble experimentally observed receptive fields. Thus, we argue, the critical objective of representational learning might be the recovery of subsampled signals, a hard and ubiquitous problem for long-range communication in the brain. Efficient coding emerges as a necessary byproduct (Isely et al. NIPS 2010).

back