PISCES / paper /data /negatives.tex
Nabidnur's picture
PISCES (ARIES #1): code, paper, notebooks, results and figures
c09a3f4 verified
Raw History Blame Contribute Delete
1.21 kB
\def\negrows{%
streaming split--merge (swap) moves, 1\% of $k$ & 2 & $-0.26\%$ & $-0.31\%$ & 2/5\\
fresher closing window (last 50\% of the final epoch) & 4 & $-0.41\%$ & $+0.49\%$ & 5/5\\
children inherit 35\% of the donor's mass each & 6 & $-0.41\%$ & $-0.28\%$ & 5/5\\
revival threshold schedule $0.3\to0.1$ & 6 & $-0.41\%$ & $-0.31\%$ & 5/5\\
graph-pruned assignment with label memory, self-verified & 9 & $-0.41\%$ & $-0.41\%$ & 0/5\\
coarse-to-fine growth (3 doublings) & 13 & $-0.42\%$ & $-0.17\%$ & 5/5\\
heavy-ball momentum on centroid moves ($\beta=0.3$) & 15 & $-0.42\%$ & $-0.14\%$ & 5/5\\
fewer probes early ($8\to16$) & 15 & $-0.42\%$ & $-0.33\%$ & 5/5\\
HNSW over the centroids instead of IVF & 17 & $-0.42\%$ & $-0.36\%$ & 3/5\\
fp16 IVF lists (same recall, faster scans) & 18 & $-0.42\%$ & $-0.29\%$ & 3/5\\
annealed assignment to the runner-up ($\varepsilon_0=0.01$) & 19 & $-0.48\%$ & $-0.46\%$ & 2/5\\
distribution-free half-mean split and score, $c\pm\E[\mathrm{sign}(p)\,e]$ & 21 & $-0.42\%$ & $-0.12\%$ & 5/5\\
empirical axis half-mean $c\pm\E|p|\,u_1$, score $N(\E|p|)^2$ & 21 & $-0.42\%$ & $-0.42\%$ & 2/5\\
batch-size schedule $0.25B\to1.75B$ & 21 & $-0.42\%$ & $-0.32\%$ & 4/5\\
}