Thesis detense Matteo Saponati (Donders Series 631)

28 November 2023

Promotors: Prof. dr. Martin Vinck and Prof. dr. Francesco Battaglia

A study of synaptic plasticity and predictive processes in biological and artificial networks

Predicting future events based on sensory input is a fundamental element of cognitive processing, impacting perception, decision-making, and motor control. Organisms can adjust their behavior to dynamically changing inputs and conserve energy by learning the temporal relationships among sensory stimuli. However, it remains unclear how the predictability of sensory input influences neuronal signals, adaptation mechanisms, and synaptic plasticity in neurons. Specifically, the contribution of the various synaptic plasticity mechanisms observed experimentally to predicting future events remains a mystery.
In my thesis, I investigated these inquiries using analytical techniques and experimental research, embracing a multidisciplinary approach within Physics. Machine Learning and Neuroscience. Chapters II and III introduced an original learning rule based on predictive processes and explored its significance in explaining anticipatory firing in neurons and diverse synaptic plasticity mechanisms observed experimentally. Notably, I demonstrated that neurons equipped with this predictive learning principle can learn spike sequences, shifting spikes to initial inputs, thus enabling anticipatory signaling and sequence recall in recurrent networks. Moreover, I applied this learning principle to recurrent neural networks with inhibitory interactions and examined how the interplay of competitive mechanisms and predictive plasticity enables neurons to selectively anticipate different input sequences. I also compared these results with conventional learning algorithms for recurrent neural networks, demonstrating that these learning principles facilitate swift and energy-efficient input classification.
Chapters IV and V focused on the impact of sensory stimulus predictability on the functional properties of individual cells and the role of collective neuronal dynamics in predictive processes. I first discussed research findings from my collaborators and me on how neurons in the mouse primary visual cortex (V1) represent sensory stimuli based on their predictability, as well as the distinct functions of excitatory and inhibitory cells in this process. Our results revealed the context-dependent nature of single-cell firing and the differing roles of excitatory and inhibitory cells in predicting future events. Furthermore, I explored the influence of excitation-inhibition balance on oscillatory activity using theoretical models and discussed its potential role in predictive processing.
In conclusion, my thesis work has the potential to significantly advance our comprehension of predictive processes, applicable to both biological systems and artificial neural networks."