- elliot plant
This project involves training a neural network to predict the outcome of dice rolls. Two loss functions are explored for the prediction task: one-hot categorical crossentropy and a custom distribution-based loss. The neural network is designed using TensorFlow and Keras. Check out the code.
Why did I make it?
I've always been curious about training Neural Nets, and I wanted to understand the relationship between a loss function and the accuracy of the network. I originally hypothesized that a loss function that calculated the aggregate loss across an entire training set would be train faster and more accurately. It turned out I had a lot of reading to do about the importance of differentiability of the loss function. Ultimately, a 1 hot encoding of the dice roll and a crossentropy loss function worked just fine for predicting the probability of rolling numbers on a dice.