Presentation slides about our journey making interactive question and answer bots using deep recurrent neural networks.

Software solution to throwing your own silent disco that doesn't require renting proprietary radio headphones. Uses virtual frame buffers, ffmpeg, and Janus (built on WebRTC) to duplicate an audio input and broadcast it to multiple other users on the same channel.

Hafner et al. made Dreamer, a model that learns a latent representation of reality. It does so in such a way that we can visualize what it learns, helping us build trust in our models. I played around with Ms.Pacman, Tetris, and MuJoCo simulator environments, like the one shown in the gif above (WalkerV2 I think).

This Deep Neural Network TensorFlow model plays snake on an LED matrix. It could only ever be as good as the python script which generated its training data. People can play on the led too using a keyboard. The current high score is held by the AI though. I trained a better model using the evolutionary machine learning algorithm "NeuroEvolution of Augmenting Topologies" (NEAT), but it is nowhere near as entertaining.