What a Convnet Is and Is Not - Convolution, Exactly

The finale states the honesty boundary: this is one exact forward pass with a given kernel. Training and any classifier softmax are outside this book.

highlighted = computed this step

What stayed exact

The convolution sums, ReLU clamp, and max-pool selection are all exact integer operations. The final pooled value in this toy pass is 9.

\text{conv}\rightarrow\text{ReLU}\rightarrow\text{pool}=9

What is outside this surface

The kernel is given, not fit here. If a classifier head followed this pass, its softmax would be a NAMED boundary, not an exact decimal computed in this book.

\text{kernel given; softmax named if a head is added}

What a convnet pass is and is not

This is one exact forward pass through convolution, ReLU, and max-pool. It is NOT training, NOT learning, and NOT a claim about a full classifier.

\text{one exact forward pass; deferred claims explicit}