An explanation of raw thoughts and review

Summary

This “report” synthesizes research for building an interpretable machine learning model that predicts my chess moves, learns my playing style, and serves as a “play chess against me” feature inside my portfolio. The key breakthrough enabling this project is Maia4all’s prototype matching approach, which achieves personalization from just 20 games- 250x more efficient than traditional fine tuning.

My assumed stack as a result of my semi extensive research, at least more extensive than i’ve ever done:

• Base Model: Maia-2 with skill aware attention
• Personalization: Maia4all’s prototype matching
• Model Selection: LogME for choosing optimal starting checkpoint
• Interpretability: Linear probing + attention rollout
• Hardware: My beautiful workaround, the ROCM 6.4.1 with an HSA_OVERRIDE because my 6700xt
• Deployment: Onyx export for web interface for simple and efficient deployment.

Expected outcome: 51-53% move prediction accuracy on personal games with interpretable explanations of what concepts drive predictions.

Table of Contents

1. Problem Defintion
2. Intro to Maia + Leela
3. Transfer Loading and friends
4. LogME: Practical Assessment of Pre-trained Models for Transfer Learning
5. Maia-2
6. Surprise Realization?