Research Project Summary:

Neuroscience and artificial intelligence (AI) have an intertwined history of collaboration. The goal of systems neuroscience is finding explanations for how the brain performs a wide variety of perceptual, cognitive and motor tasks, while AI attempts to develop machines with intelligent behaviour similar to the human brain.

In the last few decades, AI has been increasingly employed as a tool for neuroscience research which improved our understanding of brain functions. In particular, deep learning, that is a biologically-inspired AI approach , has been used as a model for various areas of the brain, such as the visual cortex and the brain's cerebral cortex. Furthermore, AI models could provide neuroscientists with novel hypotheses for how the same processes are controlled in the brain.

Self-supervised learning has been one of the most vibrant research directions in machine learning in recent years. In self-supervised learning, the supervisory signal is obtained from the data itself by leveraging the underlying structure in the data.  Inspired by the recent development in self-supervised deep learning architecture, and biological constraints according to anatomy and physiology of the canonical microcircuit, we seek to understand how the brain might implement self-supervised learning within cortical layers of the neocortex.

The neocortex accounts for approximately 75% of the mammalian brain where most of the learning happens and is believed to give us our unique cognitive abilities. The neocortex is formed of six layers I to VI (L1-6), involved in higher-order brain functions such as sensory perception, cognition, generation of motor commands, spatial reasoning, and language. In Addition to the six-layer structure, the neocortical regions are organised into columns. Despite this diversity in function, experimental studies have shown similarities in neocortical circuit organisation across areas and species, which suggests a common computational strategy to process multiple types of information. As such, uncovering the common strategy performed by the neocortical microcircuit has been of great interest to neuroscientists and AI researchers. In the canonical view of the neocortical microcircuit thalamic input first projects onto layer 4, next onto the pyramidal cells (PCs) in superficial layers  (2/3) and finally onto the deeper PCs in layer 5. Although this is a motif found throughout the neocortex, the functional role of this architecture has remained unclear. In this project, inspired by recent observations showing that thalamic input also targets layer-5 pyramidal cells and machine learning developments, we propose that the canonical microcircuit enables the brain to learn through self-supervision. In our model layer 5 PCs generate prediction errors by comparing sensory predictions originating from layer 2/3 PCs with the thalamic input received by layer 5 PCs.

In the first stage of the project, we aim to develop a self-supervised model that when trained in a sensorimotor task, which when perturbed generates error signals (or mismatch responses) akin to those found experimentally [1].  We then extend our theory to explain the neural data from similar experiments. To evaluate the scalability of our biological model, we will develop AI models to solve multi-modal self-supervised learning task, reinforcement learning and robotic control tasks.

[1] Jordan, Rebecca, and Georg B. Keller. "Opposing influence of top-down and bottom-up input on excitatory layer 2/3 neurons in mouse primary visual cortex." Neuron 108.6 (2020): 1194-1206.