11 Encoding and Decoding with Stochastic Neuron models


The influential book of Rieke (437) gives a broad introduction to the field of neural coding with a special focus on decoding. The LNP model, reverse correlation techniques, and application to receptive field measurements are reviewed in Simoncelli et al. (478).

Predictions of spike timings for a time-dependent input with models including spike history effects were performed by e.g., Keat et al. (253); Jolivet et al. (247) and different methods and approaches were compared in a series of international competitions (248; 245).

The first decoding attempts used time-averaged firing rates to decode information from a diverse population of neurons (171). Then the methods were made more precise in an effort to understand the temporal structure of the neural code (372; 57). In particular linear stimulus reconstruction from measured spike trains (437) have been widely applied.

Efficient decoding methods are a necessary requirement if a prosthetic arm is controlled by the spikes recorded from cortical neurons. Introducing spiking history effects (521) or inter-neuron coupling (399) helped to improve decoding accuracy, but the improvement of decoding techniques went in parallel with other technical achievements (473; 72; 71; 136; 521; 493; 281; 277; 385).

The discussion of the statistical principles of encoding and decoding in the present and the previous chapter is partly based on the treatment in Paninski et al. (380).