SafetyPage

Lab research on the web · Article · Distill

Zoom In: An Introduction to Circuits

Outline · 225 paragraphs

  1. Three Speculative Claims¶13
  2. Schwann’s Claims about Cells¶15
  3. Three Speculative Claims about Neural Networks¶22
  4. Claim 1: Features¶29
  5. Example 1: Curve Detectors¶35
  6. Argument 1: Feature Visualization¶42
  7. Argument 2: Dataset Examples¶44
  8. Argument 3: Synthetic Examples¶46
  9. Argument 4: Joint Tuning¶48
  10. Argument 5: Feature implementation (circuit-based argument)¶50
  11. Argument 6: Feature use (circuit-based argument)¶52
  12. Argument 7: Handwritten Circuits (circuit-based argument)¶54
  13. Example 2: High-Low Frequency Detectors¶58
  14. Example 3: Pose-Invariant Dog Head Detector¶64
  15. Polysemantic Neurons¶69
  16. Claim 2: Circuits¶76
  17. Circuit 1: Curve Detectors¶81
  18. Circuit 2: Oriented Dog Head Detection¶96
  19. Circuit 3: Cars in Superposition¶105
  20. Circuit Motifs¶111
  21. Claim 3: Universality¶115
  22. Interpretability as a Natural Science¶135
  23. Closing Thoughts¶142
  24. Glossary¶146
  25. Author Contributions¶161
  26. Historical Note¶165
  27. Acknowledgments¶167
  28. References¶172
  29. Updates and Corrections¶217
  30. Reuse¶219
  31. Citation¶221