Reproduce “Representations Have Trajectories: Characterizing Temporal Representation Dynamics in Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shivaprasad B. Gowda, Sparsh Mittal

The authors track layer-wise representation rank during neural network training and classify trajectories as persistent high-rank, collapse-and-recovery, or destructive representation collapse. Across 4,084 controlled synthetic runs, recovery runs reached training MSE below 10⁻³ in 82.89%, versus 35.11% for persistent-high-rank runs. This retrospective association is not causal, and rank alone does not diagnose learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.