Neural path / live computation

Draw a digit.
See every transformation.

This is not a prerecorded animation. Your pixels move through a real 62,902-parameter convolutional network while its actual feature maps, pooled tensors, dense activations, and normalized class scores render below.

Private by architecture. Your drawing never leaves this browser.

MODEL / LOADING 251 KB

62,902 parametersWorker isolated0 uploads

01 / DRAWING SURFACEMouse · touch · pen

Samples

02 / CURRENT READOUTNormalized class score

—

Predicted classWaiting for ink

0—

1—

2—

3—

4—

5—

6—

7—

8—

9—

Warm inference—

Ink mass—

Generation—

Live activation path

Pixels become evidence.

Select any stage to inspect its actual values. Cyan is positive activation; violet is negative.

SELECTED / CONVOLUTION ONE6 × 28 × 28

Open tensor and timing log

Stage	Tensor	Operation	Latest time
Normalized input	1 × 32 × 32	Center · scale · pad	— ms
Convolution one	6 × 28 × 28	6 learned 5×5 kernels	— ms
Maximum pooling	6 × 14 × 14	2×2 maximum · stride 2	— ms
Convolution two	16 × 10 × 10	60 sparse 5×5 channel connections → 16 maps	— ms
Maximum pooling	16 × 5 × 5	2×2 maximum · stride 2	— ms
Fully connected one	120	400 → 120 · scaled tanh	— ms
Fully connected two	100	120 → 100 · scaled tanh	— ms
Output logits	10	100 → 10 signed class scores	— ms
Class assignment	10	10 logits → normalized classes	— ms

Normalized class scores explain this model’s ranking for the drawing; they are not calibrated real-world probabilities.

What you are seeing

Real values, explained at human scale.

Convolution

Small learned kernels scan local neighborhoods and respond to strokes, corners, and more abstract combinations.

Pooling

Each 2×2 neighborhood keeps its maximum, reducing spatial size while preserving the strongest local evidence.

Dense reasoning

The final 400 spatial values become 120, then 100 learned features before the ten digit scores are normalized.

Model equations and pretrained weights are ported from Adam Harley’s MIT-licensed nn_vis, pinned to commit d625f8f. The modern TypeScript inference worker and visualization are implemented for this portfolio.

Model provenance & checksums