Papers of the day   All papers

Compositional Explanations of Neurons


Jesse Mu: New preprint with @jacobandreas: we generate explanations of the individual neurons inside deep neural networks by identifying *compositional logical concepts* that closely approximate neuron behavior (e.g. "water that isn't blue") (1/5)

5 replies, 473 likes

Jesse Mu: Compositional Explanations of Neurons will be an oral presentation at #NeurIPS2020!

3 replies, 195 likes

Jesse Mu: This Friday 9/11 at 12pm PDT I'm giving a talk at Deep Learning: Classics and Trends on Compositional Explanations of Neurons ( - open to the public! More info: Mailing list + zoom link:

4 replies, 80 likes

Jacob Andreas: New preprint led by Jesse Mu (@jayelmnop) on discovering compositional concepts in deep networks! You've heard of the "cat neuron" and the "sentiment neuron"; now, meet the green-and-brown-water neuron, the castle-or-surgery neuron, and the cheating-at-SNLI neuron. 1/

1 replies, 63 likes

Andrey Kurenkov 🤖: Wow, super cool new work on NN interpretability! So intuitive, yet seemingly powerful...

1 replies, 31 likes

Charles 🎉 Frye: As always, the @weights_biases Salon was a ton of fun! Next time, I'll be splitting the bill with @jayelmnop of @stanfordnlp/@StanfordAILab, author of, on an elegant method for explaining what single nodes in a DNN are doing RSVP:

1 replies, 13 likes

Jesse Mu: We can do the same for NLI! Check out the paper for more details (5/5)

0 replies, 9 likes

Charles 🎉 Frye: Great paper! Explaining neurons is hard, because we need a rich catalog of potential explanations. The idea here: use composition and logical connectives to generate combinatorially-many candidate explanations, then search that space efficiently.

0 replies, 6 likes

Connor Shorten: Compositional Explanations of Neurons 🔬 "Neurons may be more accurately characterized not just as simple detectors, but rather as operationalizing complex decision rules composed of multiple concepts."

1 replies, 6 likes

arXiv CS-CL: Compositional Explanations of Neurons

0 replies, 2 likes

Bolei Zhou: Great new work of generating compositional explanation using the semantics of units resulting from our NetDissect (

0 replies, 2 likes

Ronen Tamari: Cool progress towards understanding compositionality. In vision, NNs can learn compositional "concepts" with coherent meanings, for language they learn spurious heuristics. Interesting to think how to apply insights from vision to language.

0 replies, 1 likes


Found on Jun 26 2020 at

PDF content of a computer science paper: Compositional Explanations of Neurons