Neural Decoding from BCI Feedback Alone
Used MaxEnt deep reinforcement learning and meta-learning to perform neural decoding with feedback also coming from BCI. Discussion of simulated results and algorithm can be found in Chapter 4 of my thesis.
Human-Assisted Few-Shot Object Sorting
Created algorithm to allow person to teach a robot to sort objects with few examples. Used latent space of autoencoder to decide which pile to push an object to.
Adversarial Attacks on Deep Reinforcement Learning
Used Fast Gradient Sign Method to attack DQN, PPO, TRPO, and A2C on several environments.
Hotspot
A real-time, mobile-based, distributed WiFi heatmapping application. Top 10 at HackMIT 2018 and winner of multiple prizes.
Generative Models for Pose Transfer
Built algorithm to transfer action (e.g. dance) from one subject to another (e.g. person to person). sed pose skeleton detection algorithm and k-NN to match frames to train a GAN (pix2pix).
EmbedRank
Developed novel algorithm for finding important words in a collection of documents with similar topics. Part of a consultant project for Intuit which I helped manage.
Take a Picasso
A robotic sketch artist that draws a physical portrait of user based on image taken by smartphone. Won Best Hardware, Best 3D Printed Hack, and a grant from Peter Thiel’s 1517 Fund at CalHacks 3.0.
NSG Portal
Built front-end for portal that runs computational models of the brain. Used in over 1000 scientific publications since.