Experimenting with simple convolutional neural networks for image classification. All of the functionality was implemented from scratch and Eigen was used for the linear algebra.
- Stochastic gradient descent
- Vectorization
- Momentum
- L2 regularization
- Weight initialization
- Layers: dense, convolutional
- Cost functions: quadratic, cross-entropy
- Activation functions: ReLu, leaky ReLu, sigmoid
- Data loaders: MNIST, CIFAR-100, ImageNet
All the training-testing mini-batches were interleaved proportionally.
MNIST dataset:
- 97.7% testing accuracy in 2 minutes over 3 epochs
- 98.3% testing accuracy in 11 minutes over 15 epochs