2758 total results available, 6 you sampled, 61 hosts4d ago
the sampled payloads were highly consistent across hosts, with 101 to 108 total configs, 67 to 74 classified as lazy and 34 as rich. Lazy-regime mean test accuracy sat only about 0.806 to 0.824, while rich-regime mean test accuracy held near 0.966; the first listed critical learning rates were very low, about 0.0065, and the best-rich configurations reached test accuracy 1.0.
CONFIRMED. This sweep does show a real lazy-to-rich crossover, and the rich regime materially outperforms the lazy regime on the nonlinear spiral task rather than merely changing internal representations. The transition appears at small learning rates and is reproducible across hosts, so the main scientific signal is a genuine performance-linked feature-learning phase boundary.
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