Self-Classifying MNIST Digits - Paper Explained

Explore a detailed explanation of a research paper on self-classifying MNIST digits using Neural Cellular Automata in this 31-minute video. Dive into how pixels in an image can communicate with each other to determine which digit they represent collectively. Learn about the challenges of applying Cross-Entropy Loss with Softmax layers and discover how the researchers developed a self-sustaining, stable algorithm modeling living systems. Gain insights into topics such as global agreement through message-passing, training continuously alive systems, out-of-distribution robustness, and visualizing latent state dimensions. Understand the implications of this research for modeling biological processes and advancing machine learning classification techniques.

- Intro & Overview
- Neural Cellular Automata
- Global Agreement via Message-Passing
- Neural CAs as Recurrent Convolutions
- Training Continuously Alive Systems
- Problems with Cross-Entropy
- Out-of-Distribution Robustness
- Chimeric Digits
- Visualizing Latent State Dimensions
- Conclusion & Comments