Build a reproducible character-recognition experiment, improve one node at a time, and inspect exactly what the model computes.
Open browser diagnostics and integer backend checks in a new tab.
Inspect the installed 5×7 and 6×8 glyph sets and verification hashes.
1 · Create an experiment
Choose source glyphs and one target. The target is class 1; every other selected glyph is class 0. Each installed set uses a fixed, versioned bitmap cell within the image. The alphanumeric set is 5×7; symbols and chess are 6×8. Both classes are represented in each split.
Rotation and translation share geometric limits. Noise is generated after these transforms and is fixed for the lifetime of the dataset. Its Gaussian standard deviation is 10.2 × the intensity percentage before clipping to 0…255.
2 · Initialize the tree
Scale P sets the integer range 0…P. Tiling repeats the source image along both axes: 2 × 2 gives four logical copies; 4 × 4 gives sixteen. Columns and rows reduce alternately to one output. Odd edges use a zero second input.
Changing P, topology or dataset requires reinitialization. Each computed node has its own W₁, W₂ and B.
3 · Train in bounded passes
A pass chooses the specified fraction of nodes without replacement. Auto uses legacy uniform selection for 1×1/2×2 and layer-balanced coverage for 4×4. Visit order cycles output-first, random, input-first. Global proposals explore the legal parameter range; local proposals perturb the current values.
All candidates are scored against the complete Training split. Only a strictly lower exact objective is accepted. Ties keep the current node. Pause takes effect at a committed node boundary.
4 · Evaluate and export
Validation tracks model selection independently of candidate scoring. Restore best selects the earliest checkpoint with the best observed validation loss. Final Test is an explicit disclosure for the current frozen model; reuse is recorded.
C export represents the current model and its exact integer inference. Checkpoints preserve data, parameters, random state and the training cursor; derived caches rebuild on restore.
The node calculation
D = W₁X₁ + W₂X₂ · S = floor(D / P + ½) + B
S ≤ 0 → 00 < S ≤ P → SP < S ≤ 2P → 2P − S2P < S ≤ 3P → S − 2PS > 3P → P
Round once after summing the products. Half-way ties go toward positive infinity. Stored values may be 16-bit; products require 32-bit intermediates.
Loss and classification
ℓ₀ = max(0, 20Y − 9P) · ℓ₁ = max(0, 11P − 20Y)
Linear Margin is zero at outputs ≤ 0.45P for other glyphs and ≥ 0.55P for the target. BCE and Focal BCE use versioned precomputed integer lookup tables so CPU and GPU candidate ordering remains exact. Class balancing gives each class equal total influence. The displayed loss is normalized; the exact integer objective determines adoption. A sample is predicted as the target when 2Y ≥ P. Y/P is a normalized response, not a calibrated probability.
Reading the visualizations
Gate colors measure the four binary input corners in order 00, 01, 10, 11. Open the 16-gate key for their truth patterns and names. An exact corner match does not describe every interior input. Graded nodes can still contribute useful network behavior. Select a sample to trace its actual outputs.
Zoom 4× displays corner bits, 6× displays short gate names, and 8× or higher displays stored W₁, W₂ and B inside the nodes. Use 15×–30× for detailed parameter inspection. The node inspector always gives full details. Appearance includes keyboard-accessible desktop panel-width sliders; narrower screens use a stacked layout.
Training loss is non-increasing under an unchanged objective and strict acceptance. Accuracy and Validation loss can rise or fall. Repeated templates, quantized collisions and limited sample counts should be considered when interpreting results.
Proposal statistics and throughput
Global/local choice is a Bernoulli draw for each proposal; its realized proportion need not equal the configured percentage at a particular node. The last-node statistics distinguish requested candidates, unique scored triples, duplicates and one-step local fallbacks. At small P, a finite parameter domain or retry limit can cause a visible shortfall.
A candidate-sample is one candidate evaluated on one Training image along its affected path. Candidate-samples per second and node visits per second use the accumulated training-loop time. They exclude dataset generation, model initialization and separate benchmark startup; they are not counts of GPU threads or completed training epochs.
Frozen identities and disclosure history
Model UUID, revision, dataset and quantized-data hashes identify the experiment. Final Test records a hash of the evaluated parameters and the test-data identity. Its images remain sealed in the normal interface until explicitly disclosed; a checkpoint contains the raw experiment and is not an encrypted holdout.
Changing the loss weighting does not change predictions, but an earlier Final Test loss retains its old weighting and is labelled accordingly. No automatic Final Test reevaluation occurs. C packages include the most recent 50 disclosure records with an explicit total and omitted count; checkpoints preserve the complete session audit trail.
Keyboard and navigation
Drag Pan networkCtrl + wheel Zoom at pointer+ / − Zoom focused network0 Fit network← / → Child / parent↑ / ↓ Previous / next node or sampleEsc Close dialog