- Comparing deep v.s. shallow models
- Visualization of model parameter updates during training
- Error surface visualization
- Using encoder-decoder model for video caption generation and dialogue generation
- Using DCGAN to conditionally generate anime faces (can specify hair color, eye color)
- Exploring the limits of CycleGAN
- Training a Pong agent with policy gradient
- Training a DQN model to play Breakout
- Training an actor-critic (A2C) model to play both Pong and Breakout (performance charted below)