effisciencesresearch / ml4g-2.0

Improved version of the technical workshops for the 10-day ML4G camp on safety of AI systems

ML4G

This repo contains the instructions for the ML for Good bootcamp, for students and researchers interested in AI safety.

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 ML4Good

We draw inspiration from the redwood mlab, and ARENA, both of which focuses mainly on the ML engineering part. However there are a lot more workshops on strategy, goverance and conceptual AI safety during the ML4G.

Links to google colab versions of the notebooks

• agents: agents, agents hard, agents normal
• fsgm patch attack: fsgm patch attack
• gradcam: gradcam, gradcam hard, gradcam normal
• hyperparameters: hyperparameters
• image memory network: image memory network, image memory network normal
• induction heads: induction heads, induction heads hard, induction heads normal
• optimisation: optimisation
• optimizers: optimizers, optimizers normal
• pytorch tuto: pytorch tuto
• rl: A2C workbook complete, A2C workbook empty hard, A2C workbook empty, DQN workbook complete, DQN workbook empty hard, DQN workbook empty
• rlhf: rlhf
• tensors: tensors
• transformer: transformer arena, transformer, transformer hard, transformer normal
• transformer interp: exercise hard, exercise normal, logit lens probing atlas old, solutions
• vanilla policy gradient: vanilla policy gradient

For teachers, contributors and assistants

Before contributing to the project, you need to read the guidelines and follow the instructions in CONTRIBUTING.md.

ml4g-2.0's Issues

default device on transformers notebook

The reference transformer is not explicitly moved to the device, this causes issues when the default device does not coincide with 'cpu' (ie. on Mac)

https://github.com/EffiSciencesResearch/ML4G-2.0/blob/master/workshops/transformer/transformer.ipynb