UCLA and UCSD
The various methods that have been designed to decode neuronal activity in the cortex, fall under two main categories: linear or almost-linear (i.e., population code) and non-linear, such as maximum likelihood (ML). Linear estimators are sub-optimum but biologically plausible since they can be readily implemented in feed-forward linear networks. ML, on the other hand, is statistically optimum, but it is considered to be biologically implausible. It is unclear, in particular, how neurons could compute an estimate requiring a non-linear regression step and an a priori assumption about the shape of the neuron tuning curves.
We have found that a near-ML estimator can be implemented using lateral connections in a recurrent network. We will demonstrate how the lateral connections can embody an a priori assumption about tuning curve shapes and how the dynamic of the network can be used to perform a close approximation to the non-linear regression step. This work suggests that lateral connections in the cortex may allow cortical neurons to compute statistically optimum estimates of sensory or motor variables.