Convolutional neural network
A type of computer program designed to 'see' images by scanning them piece by piece, similar to how human eyes focus on small details to understand the whole picture.
A neural network architecture that uses sliding convolutional filters to automatically learn spatial hierarchies of features, making it the standard approach for image classification and object detection tasks.
A deep learning architecture employing translation-invariant convolutional kernels to exploit spatial locality in high-dimensional grid data, effectively reducing parameter complexity through weight sharing to optimize feature extraction for computer vision.
evolution
- 1980 · historyNeocognitron
Kunihiko Fukushima introduced a hierarchical, multilayered neural network inspired by the visual cortex, the direct precursor to CNNs.
- 1989 · historyLeNet-1
Yann LeCun and colleagues successfully applied backpropagation to a convolutional architecture to recognize handwritten zip code digits.
- 1998 · historyLeNet-5
The publication of the LeNet-5 architecture established the modern structure of convolutional layers followed by pooling layers for document recognition.
- 2012 · historyAlexNet
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton won the ImageNet challenge by a massive margin, triggering the modern deep learning revolution.
- 2014 · historyVGG and GoogLeNet
The introduction of deeper architectures with smaller filters demonstrated that increasing network depth significantly improves image classification accuracy.