This project is a small exercise to train a recurrent neural network to learn the XOR operator. This is inspired as a warm-up exercise from OpenAI that can be found here.

To train and validate the model, execute

python run.py

To regenerate the dataset, go inside the dataset folder and execute

python generate_dataset.py

The dataset is generated using generate_dataset.py. Given a sequence length and sample count, it will generate a dataset and save it as a CSV file.

I used an LSTM to input a 2-bit sequence of binary strings. It contains 512 hidden units and with input size of 1.

The optimizer used is the Adam optimizer with a learning rate of 1e-2.

After running for one epoch, it has reached a training set loss of 0.01842 with accuracy of 0.98872. Running a validation set after one epoch has resulted to a loss of 0.00000 with 1.00000 accuracy.

The LSTM learned the XOR function by looking at over 100000 samples for training and 10000 samples in the validation set for evaluation.