Convolutional Neural Networks
- Specialized NN designed for image processing and spatial data.
- Use: image classification, object detection, facial recognition, medical imaging.
- Data flow: hierarchical feature extraction, moving from low-level features (edges) to high-level features (objects).
- Structure: convolutional layers → Pooling layers → Fully connected layers → Output
- Activation function: ReLU in hidden layers, Sigmoid for binary classification, Softmax for multi-class classification.
- Loss function: MSE for regression, cross-entropy loss for classification.
- Learning: uses gradient descent, backpropagation, and filters (called kernels) to extract spatial patterns.
- Pros: great for images, less parameters than FNNs, translationally invariant.
- Cons: large labeled datasets, computationally expensive, no temporal awareness.
- LeNet-5: one of the earliest CNNs that promoted deep learning.
- Uses convolutional and pooling layers to extract features.
- Ends with fully connected layers for classification.
- Inspired modern CNNs like AlexNet, VGG, and ResNet.