Investigação in silico do mecanismo de reorganização sináptica do sono. Um algoritmo para maximizar a capacidade computacional de redes neurais esparsas

Abstract

The memories are stored in the brain by the persistent changes of the connectivity between neurons. Sleep plays an essential role in such changes. Research on sleep neurology has shown the activation of long-term synaptic plasticity. Experimental data point to a double role of sleep: the weakening of irrelevant memories and the reinforcement of more important ones. The hypothesis investigated in this thesis is that synaptic reinforcement and pruning, involved in memory consolidation, can bring advantages to artificial neural networks. This thesis aims to apply neurobiological sleep-dependent learning mechanisms to machine learning. For this, we carried a review of memory consolidation theories and the computational models that support these theories. Observing how the brain optimizes biological resources, the research followed the trend of artificial neural networks to apply concepts present in biological learning in machine learning. Then computer simulations were carried out to explore the hypothesis that the underlying mechanisms used by the brain for biological learning through sleep are capable of optimizing artificial neural network learning. The synaptic spatiality can bring advantage for resource economy without a learning decay, so we used a sparse artificial neural network to learn different datasets and then test if sleep could reduce the minimum of synapses that a system needs to learn patterns. The simulations were carried in different network sizes, such as different sparsity levels, several databases, in addition to modern frameworks and algorithms for artificial neural network learning. The results corroborate the hypothesis that sleeping reduces the number of synapsis required to a certain learning limit.