This project consists of implementing various neural networks to recognise handwritten Hiragana symbols. The dataset used is Kuzushiji-MNIST (KMNIST), containing 10 Hiragana characters with 7000 samples per class.
The following networks were developed to observe the strengths and limitations of varying neural networks for image classification tasks:
- NetLin - computes a linear function of the pixels in the image, followed by log softmax. The model can be run by the command line code: python3 kuzu_main.py --net lin
- NetFull - a full connected 2-layer network with one hidden layer, plus the output layer, using tanh activation at the hidden layer and log softmax at the output layer. Run the code by typing: python3 kuzu_main.py --net full --lr=0.055 --mom=0.55
- NetConv - a convolutional neural network with two convolutional layers plus one fully connected layer, all using ReLU activation function followed by the output layer using log softmax. Max pooling and a dropout layer were also used.
Each model was run for 10 training epoch on the test set and the final accuracy and confusion matrix were saved. Please refer to the report page for a discussion on the performance of the models and their architecture.