This is a heavily interactive web application, and JavaScript is required. Simple HTML interfaces are possible, but that is not what this is.
Post
wx-b.bsky.social
did:plc:equbp2kvd5xjacgew2kzxdpl
New research on grokking suggests that Softmax Collapse (SC) from numerical instability & Naive Loss Minimization (NLM) are key hurdles to generalization. Proposed StableMax & ⊥Grad show promise in addressing these.
https://openreview.net/forum?id=TvfkSyHZRA
2025-01-14T20:25:57.759Z