I found ARENA quite helpful for self-study AI safety related topics and it can work well in together with Deep Learning Curriculum. It offers colab choice for the exercise part, which is great since I don’t have much GPU support for my own computer.

Here is how I spent my time on various topics in ARENA and hope it can work as a reference for someone also interested in self-studying this material.

Total Hours: 42.7. Split into the following:

  • 8.7h Chapter 0: Fundamentals
    • skipped exercise 0.1 Ray Tracing since I found it too advanced to be necessary.
    • 2.3h exercise 0.2 CNN. I learnt about torch.as_strided related stuff.
    • 2.4h exercise 0.3 Resnet. I learnt a more detailed view about resnet, pytorch_lighting, batchnorm.
    • 2.0h exercise 0.4 Optimization. I learnt about details of various optimizer, Weight & Bias related usage.
    • 2.0h exercise 0.5 Backprop (skipped part 3 & 4 & 5). I learnt about details of back propagation & Autograd.
  • 8.0h Chapter 1: Transformers & Mech Interp
    • 3.0h exercise 1.1 transformer (I skipped part of sampling.). I learnt about details of transformer, sampling, training and inference sampling.
    • 5.0h exercise 1.2 mechanistic interpretability. I learnt about induction circuits, transformerLens, induction heads, hooks, reverse-engineering induction circuits. These material opened a new view for me about how to understand LLM. I find some part hard to understand though and skipped some of the exercise as I don’t want to spend too much time on this topic for now.
  • 16.0h Chapter 2: Reinforcement Learning
    • 2.5h exercise 2.1 Introduction to RL. it works like a memory refresher about some RL concepts. It is nice to check the detail of some RL environments.
    • 7.0h exercise 2.2 Deep Q Learning
    • 4.0h exercise 2.3 PPO. The Atari Breakout game result is shown here.
    • 2.5h exercise 2.4 RLHF.
  • 10.0h Chapter 3 Training at Scale
    • I spent most of my time on doing the exercise in Data Parallelism part and skipped most of the exercise in other sections.