Lab: Slide the Convolution Window

Mission

Run the local Convolution Window Lab. Move one 3 × 3 kernel across a fixed 6 × 6 binary image, preserve one complete calculation, compare two feature maps, and export a Convolution Evidence Record.

Starting state

The baseline uses the vertical kernel at output position [0,0], stride 1, and no padding. The 6 × 6 input produces a 4 × 4 feature map. The vertical and horizontal kernels are hand-designed teaching filters; no model is trained.

Guided start

1. Keep Vertical edge selected at row 0, column 0.
2. Run convolution and locate the covered 3 × 3 patch.
3. Read all nine signed products in the calculation table.
4. Add them and confirm the stored response.
5. Move only the column to 1 and rerun.

The input, kernel, stride, and rule stayed fixed. Only window position changed.

Procedure

1. Predict where the vertical kernel will produce its greatest absolute response.
2. Preserve the [0,0] patch, nine products, sum, and response.
3. Move the window by one column, then one row. Connect each response to its feature-map cell.
4. Record the vertical map's maximum absolute response and first position.
5. Change only the kernel to horizontal; keep the image fixed.
6. Record the horizontal maximum and first position.
7. Preserve your first feature claim before correcting it.
8. Complete Observation, Measurement, Decision, and Limitation.
9. Export the JSON and confirm it names ai-for-beginners-convolution-window-lab.

Success criteria

• one calculation contains nine products and the correct signed sum;
• vertical and horizontal kernels use the same input;
• both maximum responses and positions are recorded;
• the before/after claim remains visible; and
• the limitation rejects object identity, learned-model, transfer-performance, general-accuracy, and deployment claims.

Troubleshooting and boundary

Moving the window changes the highlighted calculation, not the already computed full map. If a sum is wrong, multiply matching positions before adding all nine signed values. Backend View is fictional local output, not a production log.

Use only supplied data. The Lab activates no camera, uploads no image, contacts no service, downloads no model, and makes no real classification decision.

robotnix-ai-lab :: local deterministic convolution
Select a kernel and window position, then run convolution.
No camera, trained model, or network service is active.