Direct Feedback Alignment Scales to Modern Deep Learning Tasks and Architectures

Explore a 35-minute video analysis of the paper "Direct Feedback Alignment Scales to Modern Deep Learning Tasks and Architectures." Delve into the limitations of backpropagation in deep learning and discover Direct Feedback Alignment as a biologically plausible alternative. Learn how this method can be successfully applied to modern deep architectures and challenging tasks, contrary to previous research. Follow along as the video breaks down the problem with backpropagation, explains Direct Feedback Alignment, provides intuition on why it works, and presents experimental results. Gain insights into neural view synthesis, recommender systems, geometric learning, and natural language processing applications of this technique.

- Intro & Overview
- The Problem with Backpropagation
- Direct Feedback Alignment
- My Intuition why DFA works
- Experiments