Lightning Talk: Backpropagation-Free Optimization in PyTorch - Andrii Krutsylo

This talk, presented by Andrii Krutsylo from the Polish Academy of Sciences, explores backpropagation-free optimization techniques in deep learning, particularly within the PyTorch framework. The speaker offers a concise overview of various training methods, including Difference Target Propagation, Direct Feedback Alignment, local loss or greedy layerwise training, and Forward-Forward learning. Each method is analyzed through the lens of practical implementation, focusing on the learning signals, necessary intermediate states, parameter updates, and the scalability challenges on modern hardware. The session aims to equip practitioners with a clear understanding of these alternative training approaches and actionable patterns for integrating them into their PyTorch workflows.