Miguel Aguilera
Biography
Ikerbasque Research Fellow, Caixa Junior Leader Fellow at Basque Center for Applied Mathematics, Bilbao, Spain
WEBSITE: https://maguilera.net
Miguel Aguilera is an Ikerbasque Research Fellow and a la Caixa Junior Leader Fellow at BCAM – Basque Center for Applied Mathematics. He uses methods from complex systems research and related areas (nonequilibrium statistical mechanics, information theory, machine learning, and nonlinear dynamics) to study the principles of neural information processing and adaptive behaviour for systems in closed-loop interaction with their environments.
Nonequilibrium associative memories in neuroscience and machine learning
Hopfield networks are a foundational model exemplifying how simple networks can exhibit complex behaviors reminiscent of memory storage and retrieval. Hopfield networks as a conceptual framework has profoundly impacted several disciplines, spanning neuroscience, statistical physics, and machine learning. More recently, an increasing interest in Hopfield networks as associative memories has surged within the machine learning community, due to the correspondence of Hopfield networks with attention mechanisms in transformer models, offering new avenues for exploring theoretical principles and designing innovative architectures. While Hopfield networks have traditionally been studied through the lens of equilibrium statistical physics, neural computation operates (in a thermodynamic sense) as an out-of-equilibrium, non-stationary process that changes dynamically, giving rise to entropy dissipation. In this talk, we will explore into the implications of extending Hopfield networks to the framework of nonequilibrium statistical physics, unveiling a diverse array of dynamic phenomena. This extended framework leads to applications in various domains, ranging from emergence of temporal asymmetries in neural circuits to applications to AI models like transformers employed in large language models.
