Thanks for coming!

Thanks to everyone for coming! I think ML is a really fascinating and important field, that I'd love to see a lot more mathematicians getting into, and I hope I managed to make a good case for that! Feel totally free to message me if you have any questions about that, or the talk.

I'd really appreciate hearing any feedback about what went well with the talk, and what I could do better in future talks!
Feedback form or just message me

• Slides

  • Note for anyone reading this: I skipped chapters 1e, 2c and 3 in the full talk (alas), but there’s some awesome content in there! I recommend reading it if they sound interesting:

    • 1e: What’s really going on with adversarial examples

    • 2c: How the network sees the data set

    • 3: An overview of reinforcement learning

    • Message me if you have any questions about them!

• Graphics

  • Image Kernels

  • GAN Generated Faces

  • Feature Visualisation

    • Note: From here on, the graphics are specific points in long articles, so I give short phrases to CTRL+F for to find them

    • Visualisation of every neuron

      • I recommend layer 4c

    • Visualising neuron interaction

      • Search for by jointly optimizing two

  • Building Blocks of Interpretability

    • Semantic Dictionaries

      • Search for Making sense of these

    • What does the network see?

      • Search for Semantic dictionaries give us a

    • Which bits are important?

      • Search for network detected features at that position

    • Which parts of the image matter?

      • Search for concisely summarizing the story of the neural network

    • How are decisions made?

      • Search for something we can optimize the factorization for

  • Activation Atlas

    • The Atlas

      • Search for how well does this work

    • What data matters for a class?

      • Search for which we ignored in this discussion

    • What differentiates nearby classes? (ie the baseball shark)

      • Search for Further Isolating Classes

  • Circuits

    • Curve Detectors

      • Outlining curves

      • Object Detection

• Best further reading:

  • Circuits & Interpretability

  • List of specification gaming examples

  • Reinforcement Learning from Human Preferences

• Learning more about ML

  • Getting started

    • fast.ai online course - really good for getting started with making something, and the basic ideas

  • Interpretability

    • distill.pub - A small journal with incredibly well-written ML papers

      • Feature Visualization, The Building Blocks of Interpretability, Activation Atlas, Circuits

      • These have papers, graphics & code

  • Reinforcement Learning

    • Sutton & Barto textbook – mathematical introduction, not overly deep learning focused

    • OpenAI Gym – deep learning focused, coding focused

• References + Further Reading:

  • Explanation of momentum: https://distill.pub/2017/momentum/

  • Image Kernels: https://setosa.io/ev/image-kernels/

  • https://web.archive.org/web/20200310063743/https://slatestarcodex.com/2020/01/06/a-very-unlikely-chess-game/

  • Maths + Navy Seal Copypasta: https://www.gwern.net/GPT-3#navy-seal-copypasta-parodies

  • AI Dungeon - play an RPG with GPT-2 https://aidungeon.io/

  • GPT-3 paper https://arxiv.org/abs/2005.14165

  • Adversarial Examples are Features not Bugs https://arxiv.org/pdf/1905.02175.pdf

    • Commentary https://distill.pub/2019/advex-bugs-discussion

  • Feature visualisation https://distill.pub/2017/feature-visualization/

    • Visualisation of every neuron https://distill.pub/2017/feature-visualization/appendix/

  • Building Blocks of Interpetability https://distill.pub/2018/building-blocks/

  • Activation Atlas https://distill.pub/2019/activation-atlas/

  • Circuits https://distill.pub/2020/circuits/

  • Meaning of neurons in early layers https://distill.pub/2020/circuits/early-vision/

  • Curve detectors https://distill.pub/2020/circuits/curve-detectors/

  • Introduction to Reinforcement Learning (fairly mathsy!) https://lilianweng.github.io/lil-log/2018/02/19/a-long-peek-into-reinforcement-learning.html

  • Examples of bad metrics https://danielmiessler.com/blog/how-to-create-bad-metrics-incentivize-wrong-behaviors/

  • Specification gaming https://deepmind.com/blog/article/Specification-gaming-the-flip-side-of-AI-ingenuity

  • List of specification gaming examples http://tinyurl.com/specification-gaming

  • Deep Reinforcement Learning from Human Preferences https://openai.com/blog/deep-reinforcement-learning-from-human-preferences/