You cannot register for this webinar

Presenting Author: Arna Ghosh, McGill University Some of the most remarkable breakthroughs are emerging at the intersection of neuroscience and AI. This convergence has been fuelled by the achievements of deep neural networks in AI and the advancements in large-scale recording techniques in computational neuroscience: Neural networks provide a testbed for exploring theories of learning in the brain, while neuroscience findings inspire the design of more sophisticated AI models. I will discuss how observations from computational neuroscience served as inspiration for developing a metric to assess the quality of learned representations in unsupervised representation learning, specifically within self-supervised learning. This metric was then used to characterize learning trajectory, and relate it to implicit bias imposed by the architectural components and the learning rule of the system. I will lay out practical recommendations for designing sample and compute-efficient self-supervised learning pipelines. Our work contributes to the ongoing synergy between neuroscience and AI, illustrating how advances in neuroscience can help design the next generation of learning systems. Bio: Arna Ghosh recently completed his PhD at McGill University & Mila - Quebec AI Institute, Canada, working with Dr. Blake Richards on neuroscience-inspired AI, specifically, representation learning and credit assignment. During his PhD, he worked on characterizing the manifold geometry of learned representations and corresponding learning dynamics in self-supervised deep neural networks of vision. Arna has been a Student Researcher at Google, and a Research Scientist Intern at CTRL Labs and Meta AI. Arna also obtained a MSc in Neuroscience at McGill University and a BTech in Electrical Engineering from the Indian Institute of Technology (IIT) Kharagpur. Arna's work has been published in top-tier AI conferences, including NeurIPS and ICLR. He was awarded the prestigious Vanier scholarship.