English  |  Français 

Efficient coding in balanced spiking networks

Principal investigator(s): Sophie Denève, Christian Machens

Cortical responses are hugely variable and heterogeneous. Moreover, the brain is amazingly robust, able to withstand lesions and neural deaths with minimal impact on behavior.  But does that really matter, given than any stimulus recruits millions of cells?  According to traditional population coding approaches, many unreliable units vote for their preferred stimulus,  a rather "brut force" approach to achieve accuracy and robustness. However, that does appear to fit with emerging data on exquisitely tuned microcircuits and synaptic plasticity rule dependent on precise spike timing. 

We thus turn the table on population coding, and consider that spiking neural networks represent (collectively) their inputs as efficiently and robustly as possible.  The excitatory-inhibitory balance ubiquitous in cortical circuits is a signature of such efficiency and robustness.  Meanwhile, neural variability is not caused by neural noise or compensated by redundancy, but due to degeneracy, e.g. multiple patterns of responses can code for the same stimulus.  Tuning curves and neural response properties in general represent solutions to an optimization problem solved at the population level. As a result, these are not fixed, but automatically changing with context and behavioral demands.