Differentiable network automaton for gradient descent optimization

• Grant number: 24/02727-0
• Support Opportunities: Scholarships in Brazil - Master
• Start date: August 01, 2024
• End date: July 01, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Odemir Martinez Bruno
• Grantee: Lucas Caldeira de Oliveira
• Host Institution: Instituto de Física de São Carlos (IFSC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 18/22214-6 - Towards a convergence of technologies: from sensing and biosensing to information visualization and machine learning for data analysis in clinical diagnosis, AP.TEM
• Associated scholarship(s): 24/16345-1 - Impact of varying neighbourhood size in LLNA's rule space exploration by stochastic gradient descent, BE.EP.MS

Abstract

The theory of Complex Networks enables the study and understanding of non-linear phenomena from varied domains such as biochemistry, social interactions and the Internet. Additionally, Network Automata are formalisms that adapts the logic of a cellular automaton for the non-regular tessellation of networks, having been used efficiently for the characterization of complex networks through the generation of temporal evolution patterns, depending on the rules of the automaton and the network topology. Since the optimization of an automata-based network classifier is a task with high computational cost, a common strategy adopted in the literature has been to limit the parameter space of the automaton and the feature extractor, followed by an exhaustive search with local optimization of the classifier, which can be seen as the grid search method. This work proposes the adaptation of the network automata and feature extractor algorithms to differentiable functions. This way, it is possible to use them as layers of a neural network and optimize all parameters through gradient descent.