MIT Researchers Present Neural Thickets Work at ICML
Source Posts on X
Cluster sourcesPhillip Isola@phillip_isolaTECH
It resolves a puzzle: if I ask "how similar are these two images," your answer will change if I add "based on color" vs "based on shape." There's no single kernel that describes both! New view: convergence not to a single rep but rather to a "thicket" of reps.
Phillip Isola@phillip_isolaTECH
At ICML, I'll present our Neural Thickets work today: 2:30–4:15 PM KST Hall A #2602 Come by if you have questions or want to chat! https://x.com/phillip_isola/status/2074632225468129284/photo/1
Phillip Isola@phillip_isolaTECH
We have a few new updates since arxiving it: 1. If you run RandOpt repeatedly, performance keeps going up 2. You can get good performance without test-time ensembling (using #1)