Understanding Neural Networks by Playing Games

Playing the Chrome T-Rex involves generating training data by actually playing the game in Chrome offline mode. We need roughly 300-500 images for training and this can be easily generated by playing the game just 5 times. We then work through the basics of Neural Networks and CNNs and then actually training and testing a model. This model can then be deployed easily within the browser and participants can see the model playing the game (with reasonable accuracy) - all within a manner of an hour. We then take this one step further by trying to understand what the model and how it correlates with how they played the game in the first place. This helps to build an understanding of how humans think and what the model actually learns and looks for in the image.