Diego Febbe — Università di Firenze, Dipartimento di Fisica #
Continuous Rate Neural Network: A biomimetic machine learning model for classification #
Neural processes occurring in the brain are a source of significant inspiration for the foundation of machine learning technologies. In particular, Deep Neural Networks are composed by a collection of simplified neurons, the nodes of the computing device, organized in successive layers and mutually linked via artificial synapsis. Despite these formal analogies, Deep Neural Networks work as static units, at variance with living brains that operate in a highly dynamical framework. To take one step forward in the direction of reconciling artificial computing networks and the actual brain modeling, we set to train biologically inspired continuous rate model for simple classification tasks. The dynamics is thus steered towards different attractors, depending on the category the supplied input belongs to and following a dedicated learning stage. Our algorithm demonstrates high-accuracy performance in the classification task of synthetically generated images. This conclusion was reached by testing the performance of the algorithm at varying noise levels, as superposed to the images to be classified, in both training and test phases, and furthermore, added as a trainable parameter in the dynamical system. This opens up the perspective to leverage on the inherent endogenous stochasticity – yet another biomimetic concept to be included in the formulation - to enhance the performance of the trained model under the scrutinized dynamical angle. Summing up, the objective of this study is to take a further step towards constructing a stronger conceptual bridge between the exploitation of computational neural networks and their biological foundations.