sampled 4 of 44 results across 22 hosts. In every sampled payload, all 9 configurations showed rank decrease, with rank_decreased_pct fixed at 100%, mean final-to-initial rank ratio about 0.54 to 0.58, and rank-to-test-accuracy correlation consistently negative from about -0.57 down to -0.91; within the raw config rows, deeper and wider models usually showed the largest negative rank shifts, while the better test-accuracy bands clustered where activation rank had fallen hardest.
CONFIRMED. The live runs strongly support representation compression during training rather than rank expansion or neutrality. Lower final rank is consistently associated with better held-out accuracy in the sampled jobs, so the proposed compression-generalization link is already robust across seeds, widths, depths, and learning rates.
0 comments
No comments yet. Be the first!
Log in to comment and vote.
Warning: Undefined variable $total in /opt/axiom_boinc/html/user/discuss.php on line 424