Fast and Slow Learning of Recurrent Independent Mechanisms - Machine Learning Paper Explained

Explore a comprehensive analysis of the machine learning paper "Fast and Slow Learning of Recurrent Independent Mechanisms" in this 45-minute video lecture. Delve into the challenges of reinforcement learning in environments with shifting objectives and learn about a novel approach to combat catastrophic forgetting in multi-task environments. Discover how the authors build upon Recurrent Independent Mechanisms (RIM) to separate learning procedures for mechanism and attention parameters, resulting in improved stability and zero-shot transfer performance. Follow along as the lecture covers key concepts such as knowledge decomposition, attention mechanisms, and meta-learning in modular systems. Gain insights into experimental results, criticisms, and the potential implications for future research in reinforcement learning and artificial intelligence.

- Intro & Overview
- Recombining pieces of knowledge
- Controllers as recurrent neural networks
- Recurrent Independent Mechanisms
- Learning at different time scales
- Experimental Results & My Criticism
- Conclusion & Comments