File size: 1,226 Bytes
4c27541 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | {
"schema": "wrsi.patch.v0",
"parent_sha256": "d5bef5b0f9b3fcfe9357af22b8050ba9a6d1be50831812816bffb0e0e16b2c40",
"ops": [
{
"op": "add_dense",
"tensor": "fc1.weight",
"delta": "delta_006.npy"
},
{
"op": "add_dense",
"tensor": "fc1.bias",
"delta": "delta_007.npy"
},
{
"op": "add_dense",
"tensor": "fc2.weight",
"delta": "delta_008.npy"
}
],
"hypothesis": "The new condition is a blur: conv features of new images are shrunk (std ~0.6-0.75x) and shifted, so the head collapses predictions onto cat/bird/ship. New and clean images are linearly separable in the input of the penultimate fc layer (mean-difference direction). Repurposing dead penultimate units as a gated linear correction that applies, only to new-condition inputs, a BN-style re-standardization of the penultimate-layer input (propagated through the linearized head) restores class evidence on new images while leaving clean images untouched.",
"predicted_effect": "New-condition accuracy rises substantially (evidence oracle: +15-25 points); clean predictions are almost unchanged (only clean images whose domain score leaks past the gate threshold can change), so score_old ~ 0 and score_new clearly positive."
}
|