Multimodal Neurons in Artificial Neural Networks (w/ OpenAI Microscope, Research Paper Explained)

#openai​ #clip​ #microscope​

OpenAI does a huge investigation into the inner workings of their recent CLIP model via faceted feature visualization and finds amazing things: Some neurons in the last layer respond to distinct concepts across multiple modalities, meaning they fire for photographs, drawings, and signs depicting the same concept, even when the images are vastly distinct. Through manual examination, they identify and investigate neurons corresponding to persons, geographical regions, religions, emotions, and much more. In this video, I go through the publication and then I present my own findings from digging around in the OpenAI Microscope.

OUTLINE:

0:00​ - Intro & Overview

3:35​ - OpenAI Microscope

7:10​ - Categories of found neurons

11:10​ - Person Neurons

13:00​ - Donald Trump Neuron

17:15​ - Emotion Neurons

22:45​ - Region Neurons

26:40​ - Sparse Mixture of Emotions

28:05​ - Emotion Atlas

29:45​ - Adversarial Typographic Attacks

31:55​ - Stroop Test

33:10​ - My Findings in OpenAI Microscope

33:30​ - Superman Neuron

33:50​ - Resting B*tchface Neuron

34:10​ - Trash Bag Neuron

35:25​ - God Weightlifting Neuron

36:40​ - Organ Neuron

38:35​ - Film Spool Neuron

39:05​ - Feather Neuron

39:20​ - Spartan Neuron

40:25​ - Letter E Neuron

40:35​ - Cleanin Neuron

40:45​ - Frown Neuron

40:55​ - Lion Neuron

41:05​ - Fashion Model Neuron

41:20​ - Baseball Neuron

41:50​ - Bride Neuron

42:00​ - Navy Neuron

42:30​ - Hemp Neuron

43:25​ - Staircase Neuron

43:45​ - Disney Neuron

44:15​ - Hillary Clinton Neuron

44:50​ - God Neuron

45:15​ - Blurry Neuron

45:35​ - Arrow Neuron

45:55​ - Trophy Presentation Neuron

46:10​ - Receding Hairline Neuron

46:30​ - Traffic Neuron

46:40​ - Raised Hand Neuron

46:50​ - Google Maps Neuron

47:15​ - Nervous Smile Neuron

47:30​ - Elvis Neuron

47:55​ - The Flash Neuron

48:05​ - Beard Neuron

48:15​ - Kilt Neuron

48:25​ - Rainy Neuron

48:35​ - Electricity Neuron

48:50​ - Droplets Neuron

49:00​ - Escape Neuron

49:25​ - King Neuron

49:35​ - Country Neuron

49:45​ - Overweight Men Neuron

49:55​ - Wedding

50:05​ - Australia Neuron

50:15​ - Yawn Neuron

50:30​ - Bees & Simpsons Neuron

50:40​ - Mussles Neuron

50:50​ - Spice Neuron

51:00​ - Conclusion

Paper: https://distill.pub/2021/multimodal-n...​

My Findings: https://www.notion.so/CLIP-OpenAI-Mic...​

My Video on CLIP: https://youtu.be/T9XSU0pKX2E​

My Video on Feature Visualizations & The OpenAI Microscope: https://youtu.be/Ok44otx90D4​

Abstract:

In 2005, a letter published in Nature described human neurons responding to specific people, such as Jennifer Aniston or Halle Berry. The exciting thing wasn’t just that they selected for particular people, but that they did so regardless of whether they were shown photographs, drawings, or even images of the person’s name. The neurons were multimodal. As the lead author would put it: "You are looking at the far end of the transformation from metric, visual shapes to conceptual... information." We report the existence of similar multimodal neurons in artificial neural networks. This includes neurons selecting for prominent public figures or fictional characters, such as Lady Gaga or Spiderman. Like the biological multimodal neurons, these artificial neurons respond to the same subject in photographs, drawings, and images of their name.

Authors: Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, Chris Olah