Memory
Actual Sihti pieces from other CIFAR training birds. The hidden target cannot be consulted.
DoDis / memory + consistency + purpose
What was there.
What fits.
What the model wants.
An image-repair experiment where the content comes from remembered images, spatial relationships come from other real images, and a separately distilled CNN graph contributes an optional preference. The three are deliberately kept apart.
Binding
Learned cross-scale and neighboring-piece co-occurrences across a separate image collection.
CNN graph
A white-box surrogate fitted to the frozen CNN on another training split. It can raise the margin without making the reconstruction correct.
Loading an experimental receipt. No real result is displayed until one exists; this page does not invent scores.
Hidden truth
Used only to score the final repair, never to choose it.
50% masked
Visible multiscale pieces are held fixed.
Reconstruction
Candidate contents come only from other images.
64 Sihti pieces
Teal: observed. Sand: held out.
—Pixel MSE (lower is better)
—Actual CNN margin on repair
—Margin gain vs corrupted input
—Accepted asynchronous piece updates
Thinking clock · descent receipt
Every candidate for each hidden piece is scored. The accepted update cannot increase the defined discrete energy. This property does not guarantee better reconstruction.
Held-out comparison
| Condition | Mean pixel MSE | Mean true-CNN margin | Reconstruction wins vs memory-only |
|---|---|---|---|
| No experimental results loaded | |||
The actual experiment
For each bird never used in the memory or graph training, hide half its 64 octave/tile residues. Retrieve candidate patches using the visible pieces only. Discrete coordinate descent compares identical candidate pools and starts under four energies: remembered content plus boundary smoothness; plus learned binding; a shuffled-binding control; and plus an optional teacher-distilled purpose term.
The key measurement is image reconstruction error against the withheld original. CNN margin is a separate outcome, measured after settling. A higher bird score accompanied by a worse image is the adversarial pull we want to expose—not a successful repair.
The score called “pixel MSE” measures the reconstructed whole image, which differs from truth only through missing multiscale residues; it is not a conventional contiguous missing-pixel hole. All models in this experiment operate over the same 64-piece Sihti decomposition. The optional CNN surrogate is trained across multiple images; it is not v3’s one-photo switch polynomial.
Research instrument, not a production image-completion model. Inspect the source and locked protocol. The viewer reads the measured v1
docs/receipt.json and, when published, the independent v2 docs/receipt_v2.json. Mouseover or graph confidence alone cannot supply missing image content.