PyData Pittsburgh - June 2024

This deck contains slides for the introductory and closing remarks from PyData Pittsburgh's June 2024 event, Radically Improving Neural Networks with Insights from Modern Neuroscience.

It also contains the slides for Patrick's lightning talk summarizing Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, a research paper from Anthropic.

Transcript

1. Sign up for the mailing list! https://news.pypgh.org
2. Welcome to all of our new members!
3. Thank you to the Swartz Center!
4. Take some pictures! 📸
5. Bonus Lightning Talk ⚡
6. We're starting to use modern language models ("LLMs") for all sorts of things!
7. ...but we don't really understand how they work.
9. the capital of France is
10. the capital of France is Paris
12. def add(x, y)
13. def add(x, y) :
14. But how does it do that, really?
15. "interpretability"
16. Scaling Monosemanticity: Extracting Intrepretable Features from Claude 3 Sonnet
17. Caveats:
18. Caveats: 1. None of this is my own work
19. Caveats: 1. None of this is my own work 2. I am not a machine learning interpretability expert
20. the capital of France is Paris
21. the capital of France is Paris 0.1 1.9 -0.5 ... 2.2
22. def add(x, y) :
23. def add(x, y) : -3.1 0.1 1.7 ... -0.8
24. Two hypotheses:
25. Two hypotheses: 1. Everything the model "knows" about the input is contained in these numbers
26. Two hypotheses: 1. Everything the model "knows" about the input is contained in these numbers 2. Combinations of neurons activating together correspond to real-world concepts
27. ...can we identify and extract combinations of neurons that correspond to relevant concepts?
28. def add(x, y) : -3.1 0.1 1.7 ... -0.8
29. def add(x, y) : -3.1 0.1 1.7 ... -0.8 0 0 0 1.4 0 0 0 0.8 0 ... 0 0 -2.7 0
30. def add(x, y) : -3.1 0.1 1.7 ... -0.8 -2.9 0.3 1.5 ... -0.5 0 0 0 1.4 0 0 0 0.8 0 ... 0 0 -2.7 0
31. def add(x, y) : -3.1 0.1 1.7 ... -0.8 -2.9 0.3 1.5 ... -0.5 0 0 0 1.4 0 0 0 0.8 0 ... 0 0 -2.7 0 (these should be close)
32. def add(x, y) : -3.1 0.1 1.7 ... -0.8 -2.9 0.3 1.5 ... -0.5 0 0 0 1.4 0 0 0 0.8 0 ... 0 0 -2.7 0 (these should be close) (this should be sparse)
33. def add(x, y) : -2.9 0.3 1.5 ... -0.5 -3.1 0.1 1.7 ... -0.8 (these should be close) 0 0 0 1.4 0 0 0 0.8 0 ... 0 0 -2.7 0 "features"
34. "sparse autoencoder"
35. ...are these "features" human-interpretable?
36. Let's find and inspect examples of model inputs that cause certain features to activate
42. What about code?
44. They only work on English-language text, right?
46. ...do these features influence the model's behavior?
47. "feature clamping"
48. 0 0 0 10 0 0 0 0.8 0 ... 0 0 -2.7 0 "features"
49. -3.1 0.1 1.7 ... -0.8 0 0 0 10 0 0 0 0.8 0 ... 0 0 -2.7 0 "features"
52. What about... dangerous features?
56. ...there are lots more features about dangerous or harmful topics ("safety-relevant")
57. So... what does all this mean?
58. This feels like a genuine step forward for understanding how modern language models actually work
59. The existence of features about dangerous or harmful topics isn’t a cause for concern
60. We're just scratching the surface!
61. The full publication is worth a read
62. Now for the main event… Radically improving neural networks with insights from modern neuroscience
63. Closing thoughts
64. You can now financially support PyData Pittsburgh
65. You can now financially support PyData Pittsburgh Contributions are tax-deductbile!
66. You can now financially support PyData Pittsburgh Contributions are tax-deductbile! Eligible for employer match!
67. You can now financially support PyData Pittsburgh Contributions are tax-deductbile! Eligible for employer match! https://pypgh.org/donate
68. Consider submitting an idea to the PyData Pittsburgh Call for Proposals!
69. Consider submitting an idea to the PyData Pittsburgh Call for Proposals! https://pypgh.org/speak
70. Next Event...
71. Data science for improving outcomes in mental health crises Pim Welle Chief Data Scientist Allegheny County Department of Human Services
72. If you’re interested in talking about building custom GPU rigs, find Robert afterward!
73. Now, let's keep the conversation going!