Skip to content

Latest commit

 

History

History
27 lines (19 loc) · 756 Bytes

README.md

File metadata and controls

27 lines (19 loc) · 756 Bytes

Bandit

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.

Features

  • 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

Performance

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

Images

Screenshot