Computational neuroscience

Computational neuroscience provides a way to bridge from the anatomical and neurophysiological properties of single neurons via their neuronal network connectivity into hypotheses about how the brain operates to perform its functions. Key issues that follow and are described here are how biologically plausible neuronal networks operate, and how information is encoded by neuronal activity, which are fundamental building blocks of how a neural system, including the brain, performs its computations. The encoding of information in the primate inferior temporal visual cortex, hippocampus, orbitofrontal cortex, and insula is described. All these areas investigated have sparse distributed graded firing rate representations. The firing rate probability distribution is close to exponential. The information increases approximately linearly with the number of neurons. Consistent with this relative independence, there is little extra information that is available because of stimulus-dependent synchrony and little redundancy. The code can be read very fast, in 20–50 ms, by dot product, biologically plausible decoding. The advantages of this code include high capacity, generalization, graceful degradation, and rapid readout of the information by biologically plausible dot product decoding. None of these are properties of local or “grandmother-cell” representations. Consistent evidence is becoming available for humans. Thus the evidence indicates that information is encoded in many cortical areas by sparse distributed graded firing rate representations, and this type of representation has major computational advantages for the brain including for language that are not met by local or “grandmother-cell” representation