Learning lesson 07 · Compose local evidence
01 · Smallest useful mechanism
A convolution selects a small receptive field, multiplies it element by element with a kernel, and sums the products into one activation. Sliding the same kernel across the image detects the same pattern at many positions. A trained convolutional network learns useful kernels and composes their activation maps into a prediction.
A shared kernel turns local pixel patches into feature maps; deeper stages combine those maps into class evidence.
02 · Experiment
Exact kernel arithmetic · local trained digit model
The convolution trace uses declared values. The digit classifier runs packaged trained weights locally and exposes activation maps without a remote model.
03 · Reason from evidence
Your brief: Build an evidence plan that treats activation maps as clues, not explanations, and confidence as a value to calibrate, not certainty.
Evidence your answer should use
There is no reveal button here. A strong response names the claim, the evidence that would test it, and what the evidence cannot establish.
Your learning artifact
Stored only in this browser. No account required; course reset does not delete it.
04 · Check your understanding
Next: A network can fit the examples it sees. The final lesson asks whether that success transfers beyond them.
5×5 input
Shared 3×3 kernel
3×3 activation map
Selected activation [0, 0]
Σ = 2
0 × -1
0
0 × 0
0
1 × 1
1
0 × -1
0
0 × 0
0
1 × 1
1
1 × -1
-1
1 × 0
0
1 × 1
1
The same nine kernel values are reused at every output position. Only the receptive field moves.
