| Project | Description | Link |
|---|---|---|
| Learning Rates | Compares various gradient descent optimization methods such as Adam, Adagrad, RMSProp and examines their impact on training loss and training time. Also implements cyclical learning rate policy | https://github.com/tamanna-a/learning-rates |
| Data Augmentation | Examines classic data augmentation and cutout augmentation, which randomly crops sections of images on CIFAR10 and Resnet-44 model | https://github.com/tamanna-a/classic-data-augmentation |
| Activation functions | Unpeel the layers and examine weights under various activation functions (sigmoid, tanh, relu, leaky relu). I examine their gradients, trying different intitilizations, and leaky Relu in solving the vanishing gradient problem. | https://github.com/tamanna-a/relu-activation |
| Neural net regularization | Batch normalization and drop out | https://github.com/tamanna-a/neuralnet-regularization |
| Bias-Variance tradeoff | Bias-Variance tradeoff | https://github.com/tamanna-a/bias-variance |