Convolutional Neural Network
A CNN shares a local kernel across spatial (or temporal) positions, building translation-equivariant features. It is the inductive bias that cracked modern computer vision.
Definition
A convolution layer computes (K ∗ x) at every location with a small kernel K, then stacks ReLU, pooling, and more convolutions. Weight sharing slashes parameters versus a dense net on pixels. 1-D CNNs apply the same idea to time series and text.
Why it matters
If the signal is local and shift-tolerant — images, spectrograms, order-book images — a CNN is still a strong, cheap baseline. Transformers won large-scale vision by adding more data and compute, not by making locality useless.
Case
AlexNet (2012) cut ImageNet error so hard that “deep learning” became the default research program. The architecture was convolutional; the shock was scale plus GPU plus ReLU plus dropout.
How to read it
Ask what the receptive field is at the last layer. If it cannot see the pattern you claim to detect, the net is classifying textures, not objects — or not the event you think.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Gradient-based learning applied to document recognitionLeCun, Bottou, Bengio, Haffner · 1998 · Proc. IEEELeNet — operational CNNs.
- ImageNet Classification with Deep Convolutional Neural NetworksKrizhevsky, Sutskever, Hinton · 2012 · NeurIPSAlexNet — the ImageNet shock.