An empirical investigation of biologically plausible Hebbian systems

This paper presents an investigation into three algorithms that have pattern matching and learning capabilities. There are many neural network models in use today that are computationally viable for numerous tasks. These models tend to be inspired from neuroscience; this is evident from some of the terminology used such as threshold levels. However, many of these models incorporate complex algorithms for which there is no biological basis; in addition, these models tend to be multi-pass systems and are too slow for biological implementation. Backpropagation is one such example. The algorithms discussed in this paper share the common property of being good candidates for neural implementation; they are one-pass systems incorporating variations of a learning rule that is known to exist in real neural networks. We will verify the maximal performance of these three systems. Of more importance, however, are their capabilities under biologically constrained conditions; that is, what is the system performance (1) if learning begins with non-optimal parameters, (2) if the system is only capable of low-precision mathematics, and (3) if the learning results are corrupted. These biologically motivated issues will be the focus of the investigation.