Professor Alyosha Efros
- Demonstrated complex emergent properties of multi-agent systems with deep reinforcement learning in Professor Alyosha Efros’s lab with Deepak Pathak.
- Used Pytorch and the Unity Machine Learning Agents platform to develop the project. The project involved combining graph neural networks under the PPO framework to allow for the creation of agents with dynamic morphologies.
- A video of the project can be found here: https://pathak22.github.io/modular-assemblies/
- This project won the Genetic and Evolutionary Computation Conference's Virtual Creatures Competition https://virtualcreatures.github.io/
- Accepted at NeurIPS 2019 (Spotlight presentation)
- In another project with Deepak Pathak, I created a reinforcement learning environment with Unity to demonstrate the shortcomings of current implementations of curiosity-based exploration.
- Many implementations of curiosity-based agents can get attracted to states with high variance since predicting the next state can become difficult. Thus, the environment includes a TV-like screen that randomly changes when the agent takes a certain action. This causes the agent to stop exploring the rest of the maze.
- A video of that work can be found here: pathak22.github.io/large-scale-curiosity/ and it appears in another paper as well here: https://arxiv.org/abs/1906.04161
- That work also appears in:
Professor Bin Yu
- Applied a scattering transform to early layers of Convolutional Neural Networks to increase their interpretability in Professor Bin Yu’s lab.
- Used Pytorch and Tensorflow to run experiments on multiple datasets, such as Imagenet.
- As a result of this work, I am a Center for Science of Information Channel Scholar. More information about that can be found here.