leogrin / ml4g

Machine Learning for Alignment Bootcamp

This repo contains the instructions for the ML for Good seminar proposed by EffiSciences for French speaking students and researchers interested in AI safety.

We take over the program of the Redwood Research mlab which lasts 4 weeks. We think it is possible to see the essence of the program in 10 days: 7 days of training and 3 days of projects.

The program is aimed at beginners in machine learning, but is quite ambitious, and we hope that even advanced students will enjoy participating in this program.

Curriculum ML4G

First week: Classes, tutorials, peer coding:

• w1d1 - pytorch + CNN
• w1d2 - Interpretability - optimization
• w1d3 - transformer architecture
• w1d4 - experiments on GPT-2 - huggingface
• w1d5 - RL
• w1d6 - theoretical intepretability
• w1d7 - AI safety review

Second week: Start of the project

• w2d1 - Presentation of possible projects and choice of project
• w2d2 - Literature review on the issue
• w2d3 - Beginning of the work in groups of 2

We draw inspiration from the redwood mlab, which focuses mainly on the ML engineering part. In comparison, we will spend more time on conceptual aspects.