Gradient-Based Learning Applied to Document Recognition
The paper that put convolutional networks on a firm footing: local receptive fields, shared weights, and pooling give translation-tolerant features with far fewer parameters than dense layers. It also describes training whole systems end to end with gradients, including document-level pipelines.
Key ideas
- Weight sharing encodes a prior that features are local and repeat across the image
- Pooling buys invariance at the cost of precise position
- Graph transformer networks trained multi-stage systems end to end
Why read it now
Convolution as weight sharing is the mathematical idea behind every vision backbone. Reading the original makes the parameter counting concrete.
Question to keep in mind
How many parameters does a convolutional layer have compared with a dense layer over the same input, and why does that gap matter for generalisation?