Foundations of Neural Networks and Transformers · Lesson 8
Convolutional Neural Networks
CNNs process grid data such as images. Convolution filters slide over the image to detect local patterns like edges and textures; pooling reduces resolution while keeping important features.
Stacked convolution layers build a hierarchy from edges to shapes to objects. Classic architectures include LeNet, AlexNet, VGG and ResNet.
Transfer learning reuses a CNN pre-trained on a large dataset and fine-tunes it for your task with far less data.
Key points
- Filters detect local patterns
- Pooling downsamples feature maps
- Hierarchy: edges → shapes → objects
- Transfer learning saves data and time