Modelling Learning Characteristics of a Cortical Neural Network in an Artificial Neural Simulation
In 2001 Shahaf and Marom published a paper detailing a thoughtful experiment, in which a network of
cortical neurons developing ex vivo was shown to be capable of learning “arbitrarily chosen tasks”. A series of training cycles, in which the network was stimulated until it elicited a desired response, drove the network to modify its circuitry until eventually the desired response was directly elicited by the stimulus.
From this it was concluded that the stimulus acted as a driving force for circuitry modification, and the removal of stimulation acted as a form of reward for the desired behaviour. As the first demonstration of task learning in a network of real cortical neurons and the first empirical account of learning without a separate reward mechanism the Shahaf and Marom report is truly pioneering.
However the very claim that learning can be attributed to a neural network is novel, and as yet is contended. In this project we aimed to provide the complement to the findings of the Shahaf and Marom report, firstly by defending the novel claim that learning can be attributed to a network and secondly by furthering current knowledge on the core mechanisms underlying the learning behaviour shown by the Shahaf and Marom network. To do this we attempted to reproduce the discovered learning behaviour using a minimal model of a biological network in an artificial simulation. Several models investigating a host of phenomena observed in biological neural networks shed some light on their possible roles in producing biologically realistic learning behaviour, although the behaviour itself was not wholly reproduced. Synaptic depression was identified as a key element in generating biologically realistic reverberating responses, and its correct application was identified as a matter for future investigation.