Learning to change remembered content

The question

Can a shared learned readout recover an operation-relevant difference from stored VSA memory, so a fixed pattern bank can express a useful change?

Why this experiment?

The earlier memory-only selector left useful changes unrealized. A shared learned readout tests whether recovered information can better guide the same supplied operation.

How it works

Train a shared prediction of the feature difference for a supplied exchange, while keeping the image codec, acquired patterns and operation fixed. Evaluate on separate photographs and count no-change outputs as well as gains and losses.

What happened

The trained readout improved 23 requests, worsened none and left 41 unchanged, with 23 changes across 12 photographs. Untrained improved 33, worsened 27 and left four unchanged. Post-experiment analysis found that training avoided all 27 untrained losses by making no change, but also lost ten useful changes.

Mean regional squared RGB error was 0.000263286 for Trained: 62.9% below leaving memory unchanged, but 5.5% above Untrained and 1.55% above corrected independent selection. The primary 95% source-bootstrap interval crossed zero. The learning-improvement and bounded-candidate criteria were not met.

The trained readout had no observed regressions against unchanged memory on this panel. Untrained had 25 regional error doublings; corrected independent selection had eight. Matching memory mattered: Trained beat inference using another photograph’s memory. Privileged Clean selection improved 32 requests, worsened none and left 32 unchanged.

Validation vector loss fell from 327.946 to 20.538, but an analytic zero-difference output scores 0.836 on the same saved targets. The budget ended at 1,024 steps while loss was still declining. This post-experiment check identifies unresolved optimization, not a ceiling for the readout architecture. A zero-difference output would make no changes, so vector loss and decoded usefulness remain separate tests.

The global scale learned a value of 0.2495. The residual correction’s output norm was 12.8% of the scaled input’s norm; without a scalar-only control, its contribution to useful pair choices is not isolated. Independent validation checked 1,325 files and recomputed 768 image measurements. State continuity, held work, the bank, codec and pixels outside the edited core were preserved.

Limits and next question

No-ops select the same pattern twice; they are not learned confidence judgments. Zero observed losses across these 64 requests does not establish zero risk elsewhere, and the measured tradeoff does not change the failed primary criterion.

The operation and positions were supplied. Withheld position groups appeared in validation tasks and all-position memories, although excluded from gradient and basis fitting. Evaluation photographs were independent; these were not untouched zero-shot position tasks. Clean and the exact-feature reference remain privileged.

The two transfer groups differed: mean error improved 14.6% versus Untrained for withheld pairings of trained positions and worsened 25.4% for positions excluded from gradient and basis fitting. These are descriptive subgroup results, not separate confirmation studies.

Next question. Does simple attenuation explain the conservative behavior, and can convergence-checked readout training recover more useful changes without bringing back the observed harms? No extra training, parameter sweep or automatic integration followed this result.