Responsible & Safe AI at GSFC Univ Vadodra

Bias, Machine Unlearning, Interpretability, Guardrails, Jailbreaking, and many more

1. Responsible & Safe AI 1 Aug 2024 Inaugural Ceremony GSFCU ACM Student chapter Vadodra , Gujarat Ponnurangam Kumaraguru (“PK”) # ProfGiri CS IIIT Hyderabad Vice President, ACM India TEDx Speaker https:// precog.iiit.ac.in / /in/ponguru @ponguru

2. 2

3.     3

4. Know the Audience Students / Faculty / Others Any of you have attended my talk(s) before?  4

5. What is AI? 5

6. What is Responsible & Safe? 6

7. What is Responsible & Safe AI? 7

8. 8

9. 9

10. 10

11. 11 Observations?

12. 12

13. 13

14. 14

15. 15

16. 16

17. 17

18. 18

19. 19

20. 20 https://translate.google.co.in/

21. 21 https://translate.google.co.in/

22. 22 https://translate.google.co.in/

23. 23 https://translate.google.co.in/

24. Activity Please do any prompting in any of these or other platforms, get them to give you biased response, do not do gender bias HINT: There are very nice prompts that students have come up with in the past  24

25. 25

26. 26

27. 27

28. 28 Guardrails

29. 29

30. 30 Jailbreak

31. What is an alignment problem? 31

32. What is an alignment problem? 32 https://youtu.be/yWDUzNiWPJA?si=wSDO4i_EMrHzHYDP  

33. High- level instantiation: ‘RLHF’ pipeline First step: instruction tuning! Second + third steps: maximize reward 33 https://arxiv.org/pdf/2203.02155

34. Rouge AIs We risk losing control over AIs as they become more capable. Proxy gaming: YouTube / Insta – User engagement – Mental health 34

35. What is going on?  https://www.youtube.com/watch?v=lnyuIHSaso8&t=75s 35

36. Questions? 36

37. Forms of unlearning  Exact unlearning  Approximate unlearning Unlearning via differential privacy  Empirical unlearning, where data to be unlearned are precisely known       (training examples) Empirical unlearning, where data to be unlearned are underspecified       (think “knowledge”) 37

38. Graph Unlearning What is it? 38

39. Graph Unlearning 39 Node feature unlearning Node unlearning Edge unlearning

40. 40 Interpretability

41. New techniques and paradigms for turning model weights and activations into concepts that humans can understand https://en.wikipedia.org/wiki/Neural_network https://en.wikipedia.org/wiki/Artificial_neural_network 41 Interpretability

42. Interpretability: Mechanistic Reverse-engineer neural networks Explaining neurons and connected circuits 42

43. 43

44. 44

45. 45

46. 46

47. 47

48. Bias in LLMs Current systems like ChatGPT employ guardrails, and do not respond to biased content Users on the Web leave out key contexts, which make LLMs think the content is biased This negatively affects user engagement LLMs must be able to explore and ask more questions Our work aims to make LLMs bias-aware – context resolves confusion! 48

49. Other Directions Probing   Robustness Jailbreaking …. 49 https://precog.iiit.ac.in/pages/publications.html

50. 50 https://precog.iiit.ac.in/teaching/responsible-ai-nptel-f24/index.html

51. 51 Search for: Ponnurangam Kumaraguru https://www.linkedin.com/in/ponguru/ https://twitter.com/ponguru

52. Interested in working with us? 52 Full time Research Associates  PhD Students  Interns 

53. 53 https://precog.iiit.ac.in/

54. Acknowledgements Precog members Collaborators 54