In this collaborative project, we compared the performance of a variety of machine learning techniques for classifying hand-written digits, and connected our final model to a Gradio interactable drawing pad in Google Colab. To the right is a video we made to showcase our work.

We evaluated the performance of three approaches with a variety of hyperparameters by using grid searches and Tensorflow Keras. These three approaches were based on an autoencoder, a feedforward neural network, and a convolutional neural network. As expected, the CNN performed the best, with the least overfitting and highest accuracy of the three (98.6%), though it still had some limitations due to the content of the dataset.