Improving deep neural network random initialization through neuronal rewiring

Autor(es)

Scabini, Leonardo ; De Baets, Bernard ; Bruno, Odemir M.

• Tipo de documento: Artigo Científico
• Fonte: Neurocomputing; v. 599, p. 13-pg., 2024-07-17.

Resumo

The deep learning literature is continuously updated with new architectures and training techniques. However, weight initialization is overlooked by most recent research, despite some intriguing findings regarding random weights. On the other hand, recent works have been approaching Network Science to understand the structure and dynamics of Artificial Neural Networks (ANNs) after training. Therefore, in this work, we analyze the centrality of neurons in randomly initialized networks. We show that a higher neuronal strength variance may decrease performance, while a lower neuronal strength variance usually improves it. A new method is then proposed to rewire neuronal connections according to a preferential attachment (PA) rule based on their strength, which significantly reduces the strength variance of layers initialized by common methods. In this sense, PA rewiring only reorganizes connections, while preserving the magnitude and distribution of the weights. We show through an extensive statistical analysis on image classification tasks that performance is improved in most cases, both during training and testing, when using both simple and complex architectures and learning schedules. Our results show that, aside from the magnitude, the organization of the weights is also relevant for better initialization of deep ANNs. (AU)

• Processo FAPESP: 19/07811-0 - Redes neurais artificiais e redes complexas: um estudo integrativo de propriedades topológicas e reconhecimento de padrões
• Beneficiário: Leonardo Felipe dos Santos Scabini
• Modalidade de apoio: Bolsas no Brasil - Doutorado
• Processo FAPESP: 21/09163-6 - Ciência das redes para otimização de redes neurais artificiais em visão computacional
• Beneficiário: Leonardo Felipe dos Santos Scabini
• Modalidade de apoio: Bolsas no Exterior - Estágio de Pesquisa - Doutorado
• Processo FAPESP: 21/08325-2 - Análise de autômato de rede (network automata) como modelo para processos naturais e biológicos
• Beneficiário: Odemir Martinez Bruno
• Modalidade de apoio: Auxílio à Pesquisa - Regular
• Processo FAPESP: 18/22214-6 - Rumo à convergência de tecnologias: de sensores e biossensores à visualização de informação e aprendizado de máquina para análise de dados em diagnóstico clínico
• Beneficiário: Osvaldo Novais de Oliveira Junior
• Modalidade de apoio: Auxílio à Pesquisa - Temático