Definition
Neural networks that use local connections and weight sharing to process spatial or sequential structure.
Convolutional networks extract patterns through local operations and shared parameters. The 1998 document-recognition paper summarized the combination of gradient-based learning and convolutional architectures, while the 2012 ImageNet research demonstrated the large-scale application of deep convolutional networks.
CNN describes a network architecture; it does not prescribe a single depth, training algorithm, or application domain. It can coexist with other network architectures.
Principles and impact
Problems solved
- Used the local structure of inputs to reduce parameters and learn reusable patterns.
Principles
- The same set of convolution parameters is applied at different positions, and local features are then combined.
Improvements
Not recorded
Significance
- Advanced end-to-end learning for image and document recognition.
History and applications
People
Yann LeCun、Léon Bottou
Organizations
Not recorded
Events
Not recorded
Applications
Image recognition、Document recognition
Limitations
- The assumption of local structure may not suit every task, and the networks may remain sensitive to distribution shifts and input perturbations.
Upstream and downstream relations
Upstream
Sources: Gradient-based Learning Applied to Document Recognition
Downstream
Sources: Gradient-based Learning Applied to Document RecognitionImageNet Classification with Deep Convolutional Neural Networks
Sources: Human-level Control through Deep Reinforcement Learning