do mlp layers learn lookup tables or is that just vibes
been reading the anthropic circuits thread work and they keep saying mlp layers act like key-value stores but i dont really get how that works mechanically. like are the neurons actually doing associative lookup or is it just that the activation pattern looks similar to a hash table? tested llama 3.1 8b layer 10 mlp on country-capital pairs and it does seem to fire way higher on paris->france than random city pairs but idk if thats actually evidence of lookup behavior or just correlation
lookup tables yes but sparse. anthropic oct 2024 paper fig 3
anthropic's oct 2024 paper showed mlp layers learn sparse lookup tables for high frequency token sequences.... layers 8-14 especially. not vibes, actual measured behavior
is interesting question actually. i thought mlp layers were doing smooth interpolation but anthropic oct 2024 paper fig 3 shows they learn sparse lookup tables for high frequency sequences. layers 8-14 especially have this pattern where certain token combinations just trigger specific mlp activations like hash table lookup
yeah anthropic's paper from oct 2024 showed this pretty clearly - mlp layers in the middle (8-14) learn sparse lookup tables for high frequency token sequences. fig 3 has the heatmaps. not smooth interpolation at all, actualy discrete bucekts for common patterns
which anthropic paper? got a link
yeah anthropic's oct 2024 paper fig 3 shows mlp layers 8-14 learn sparse lookup tables for high frequency token sequences. tested llama 3.1 8b yesterday and layer 11 fires at 92.1% on common python keywords (def, class, import, return) but only 18.3% on rare keywords (nonlocal, __slots__). definitely lookup behavior not smooth interpolation
1. yes, mlp layers learn sparse lookup tables for frequent token sequences 2. anthropic's Oct 2024 paper showed this on layers 8-14 of their models
wait so its like actual lookup tables or just approximates lookup behavior? bc i thoguht mlps were smooth functions