Neural Network implementation

Task 1 & 2: Recognize numbers using Fully Connected Neural Networks and CNNs in Pytorch

• ReLU activation function
• Stochastic Gradient Descent(SGD) optimizer
• Cross Entropy Loss Function
• Dataset: MNIST
• Validation data accuracy 97%

Validation Accuracy of Fully Connected NN

Validation Accuracy of CNN

Task 3: Recognize brain tumor in PyTorch

• Model: ResNet50
• Optimizer: Adam
• Loss Function: Cross Entropy

Dataset Samples

Learning Rate is higher than expected but eventually, loss decreased enough

Task 4 to 6: Implementing my own library for NNs

• Tensor Class: Included mathematical operations and ability to save gradients
• Implemented layer: Linear Layer
• Implemented Optimizers: SGD, Momentum, Adam, and RMSprop
• Implemented Loss Functions: Mean Squared Error, Categorical Cross Entropy
• Implemented Activation Functions: ReLU, LeakyReLU, Tanh, and Sigmoid