MODELIMAGE / EXPERIMENT 02

The world learns
its missing pieces.

Two identical visual worlds, two different experiences. Neither learner is told which fragments form a face. It discovers recurring parts and tests possible relationships. Hide a fragment and inspect what the past makes it expect.

NEW → V3 multiscale image model
← Original tile assembler
Open source and test gates ↗
All processing happens inside your browser.

Teach two different histories

A and B see the same kinds of fragments and the same distractors. The only difference is a three-part rule. Training never supplies the rule or the names of the fragments.

World A experiences0
World B experiences0
Distinct atoms A0
Distinct atoms B0

Learned movable fragments

The dictionary below is learned from raw 16×16 RGB patches. Each codeword can be recognized in any position. It is not told the location or name of an eye, mouth, or distractor.

Visible learning trace

No experience. Only sensor pixels and a fixed capacity to form visual atoms and test local associations.

Test what the system believes

Select a new world, then cover just the third fragment or hide half the image. Both learners receive the exact same observed pixels. The missing area is never available to their prediction code.

What both learners see · holes masked
Actual scene · only you can see it
Pairwise-only completion (one context part)
World A learned two-context completion
World B learned two-context completion
Learned atoms alone · readable fragment vocabulary

Which fragments explained the missing one?

The outlined sensor fragments are the evidence selected by the strongest learned relation for A's missing piece.
Training first makes the test meaningful. For each learned world we measure the exact missing fragment code on unseen scenes against a one-context control. We also measure actual hidden-region pixel MSE.
How much is learned, and what are the limits?
  1. Unsupervised atoms. Each patch is reduced to 4×4 RGB averages. Similar appearances form an online competitive codebook, and each codeword retains the average of its full-resolution 16×16 training pixels for generation. This is a readable patch dictionary, not a deep neural network.
  2. Self-discovered local relationships. From complete training images, every spatial site tries all pairs of its eight neighbor offsets. It accumulates conditional counts and scores relations by their predictive lift over the frequency of the target atom, discounted by support. No privileged eye/eye/mouth offsets or product feature are inserted into the learner.
  3. Completion. Visible patches are encoded once; hidden patches are replaced with learned codewords chosen using visible context, with the strongest relationships filled first. Predicted pixels come from the learned prototypes—not from the hidden sensor image or from an original tile supplied at test time.
  4. Limits. This synthetic world deliberately reuses a small alphabet of 16×16 glyphs. Discovering an association here does not imply learning natural-image semantics. This model is still single-scale, local and discrete. Half-scene completions can be wrong or generic, and it has no mechanism yet for generating previously unseen detailed textures.
  5. Controls. Two-input versus one-input inference on exactly the same learned observations; opposite worlds with identical individual and pairwise statistics; erased learners; held-out unseen seeds; a masked-pixel nonleakage test in CI. The geometry search has a fixed neighborhood radius, but the successful offset pair is selected from data.

Next: progressively finer residues, deformable group hypotheses, online Hebbian competition and tested real-image transfer. This page is a controlled first step toward a comprehensible generative image model, not a claim of qualia or new image-model state of the art.