Compensatory Hebbian learning for categorisation in simulated biological neural nets

Using a reasonably accurate fatiguing leaky integrate and fire (FLIF) neural model, and biologically plausible compensatory Hebbian learning rules, simulations categorise benchmark machine learning data. The FLIF model is a simple, efficient point model with discrete cycles roughly corresponding to 10 ms. of biological time. The model is applied to the yeast categorisation task and the results are compared with those of other mature machine learning algorithms, including a new Kohonen net. Synaptic weights are changed following a compensatory Hebbian rule that includes the total synaptic weight of a neuron. The neural model leads to spontaneous neural firing that enables neurons not directly stimulated by the environment to be included in the neural categorisation circuit. The network is sparsely connected, and broken into two subnets, with the first subnet directly stimulated by the environment, and using compensatory learning based on the strength leaving the neuron. The second subnet initially fires only spontaneously, and uses compensatory learning based on the weight entering the neuron. After learning, new items are categorised based on a Pearson measurement comparing the firing behaviour of the second subnet on trained items, and the test item. The simulation is self-organising using only unsupervised learning. This “biologically” plausible learning mechanism and network is close to the machine learning algorithms’ performance; the biological network categorises 53% correctly, while the Kohonen net categorises 56% correctly. This neural simulation is incomplete, but supports further developments in biological neural cognitive architectures.