Abstract
Neural connectionism is a common theoretical abstraction of biological neural networks (1–3) and a basis for common artificial neural networks (4). Yet, it is clear that connectionism abstracts out much of the biological phenomena significant and necessary for many cognitive-driven behaviors, in particular intra-neuronal and inter-neuronal biochemical processes (5–8). This paper presents a model which adds an abstraction of these processes to a standard connectionism-based model. Specifically, a sub-system determines the synaptic weights. The resulting network has plastic synapses during non-learning-related behavior, in sharp contrast with most common models in which synapses are fixed outside of a learning-phase. Some synapses introduce plasticity that is causally related with behavior, while in others the plasticity randomly fluctuates, in correspondence with recent data (9,10). In this model the memory engram is distributed over the biochemical system, in addition to the synapses. The model yields better performance in memory-related tasks compared to a standard recurrent neural network trained with backpropagation.
