Sriram Somasundaram

I am an undergraduate researcher in Computer Science at USC, where I am advised by Prof. Joseph J. Lim. I am a part of USC's Cognitive Learning for Vision and Robotics Lab, and my research interests are reinforcement learning, robotics, and machine learning.

I am completing my undergrad at USC in December 2018 with a Bachelors in Computer Science. In 2015, I was a research intern at the Mellins Lab at Stanford University under Prof. Elizabeth Mellins. I've also interned at Riot Games, Raytheon, and Quid.

Email  /  CV  /  Google Scholar  /  LinkedIn


I'm interested in investigating how learning algorithms can enable machines to acquire complex skills in real-world settings. My research experiences thus far delve into learning and interpreting deep representations for complex skills and enabling machines to smoothly compose and execute skills for hierarchical tasks.

Composing Complex Skills by Learning Transition Policies with Proximity Reward Induction
Youngwoon Lee*, Shao-Hua Sun*, Sriram Somasundaram, Edward Hu, Joseph J. Lim,
International Conference on Learning Representations (ICLR), 2019
project page

Transition policies enable agents to execute learned skills smoothly to perform complex tasks.

Neural program synthesis from diverse demonstration video
Shao-Hua Sun*, Hyeonwoo Noh*, Sriram Somasundaram, Joseph J. Lim,
International Conference on Machine Learning (ICML), 2018
project page

We propose a neural program synthesizer that explicitly synthesizes underlying programs from behaviorally diverse and visually complex demonstration videos.

pH-susceptibility of HLA-DO tunes DO/DM ratios to regulate HLA-DM catalytic activity
Wei Jiang, Michael J. Strohman, Sriram Somasundaram, Sashi Ayyangar, Tieying Hou, Nan Wang, Elizabeth D. Mellins,
Scientific Reports vol. 5, p. 17333, Nov. 2015

We investigated the pH-dependence of DM-DO-mediated class II peptide exchange and identified an MHC-II allele-independent relationship between pH, DO/DM ratio, and efficient peptide exchange.

Course Projects

Voice Conversion between Male and Female Speech
Sriram Somasundaram, Andrew Szot, Md Nasir, Arshdeep Singh, 2017

Investigated voice conversion between male and female audio using multiple approaches including a CycleGAN on audio spectrograms and speaker conditioned generators with and without latent space supervision.

this guy makes a cool website