I'm an AI Engineer at Together AI, working on finetuning and distributed training of LLMs, building large scale GPU systems, and pushing the boundaries of Generative AI. I've worked in the field of generative modeling for over 8 years, largely focused on speech processing, computational music, and natural language processing. I graduated from University of California, Berkeley with a Bachelors of Science in Electrical Engineering & Computer Science in 2020, focused on Machine Learning. In the past I worked at Google, Berkeley AI Research, and Brilliant Inc. In my free time, I produce music, sketch, and game.
Due to privacy concerns, please contact me on LinkedIn for additional contact information.
Non-Degree• Sep '21 - Present
B.S. Electrical Engineering and Computer Science • Aug '16 - May '20
AI Engineer• July '23 - Current
Research Engineer• April '22 – April '23
Deep Learning Engineer• July '20 – Feb '22
Researcher & Grader • Jan '19 - Jan '20
Software Engineer Intern • May '18- Aug '18
President & Project Lead • May '17- Dec'19
Course Author• May '17- Aug'18
Ashwin Balakrishna*, Brijen Thananjeyan*, Jonathan Lee, Arsh Zahed, Felix Li, Joseph E. Gonzalez, Ken Goldberg
Conference on Robot Learning (CoRL), 2019. Oral Presentation. PDF
Ashwin Balakrishna*, Brijen Thananjeyan*, Jonathan Lee, Arsh Zahed, Felix Li, Joseph E. Gonzalez, Ken Goldberg
Real World Sequential Decision Making Workshop at the International Conference on Machine Learning (ICML), 2019. PDF
My work cetralizes around a combination of Deep Learning, Speech Processing, and Reinforcement Learning. In the past, I've worked on proejcts involving Computational Music, NLP, Text-To-Speech, Model Based Reinforcement Learning, and Dynamics Estimation. I'm fluent in TensorFlow, PyTorch, and various other ML frameworks.
Used uncertainty estimation to create an active learning framework for physics estimation. Achieved a >50% decrease in required data. Paper and code pending publication.
Inferred style-embeddings from text to improve generated speech. Second network trained from text, pitch, and rhythm embeddings. Improved F0 Frame Error by 8% with audible improvement.
Invented a generalized class of metrics over the space of policies for a given Markov Decision Process. Trained a Siammesse Network to compare tasks, with distance related to the distance between optimal policies of the two tasks. Can be used for improved exploration of policy initialization in a meta-learning setting.
Built a custom mount for a Kinect on a Sawyer robot arm. Used the depth reading to identify objects using a Gaussian-Mixture Model. The grasp angle is determined by an approximated integral optimization.
Project for TreeHacks 2018. Used NLP and Sentiment Analysis to spot biased news, and DNNs to piece together articles to create and unbiased news article.
Simultaneously and independently optimize all solutions on the Pareto front. Implement and improve the existing MO-CMA-ES algorithm to operate a Baxter robot quickly and adaptively in production for a multi-objective problem such as collision avoidance.
Designed and implemented original Phase-Functioned LSTMs with autoencoders in Python/TensorFlow to generate music wth repeated patterns using the Magenta framework.
I like to put my digital signal processing skills to a more creative use with electronic music production. I mostly produce for myself, so I haven't published much of my works yet, but here's an inside look at some of the projects I'm proud of or working on.
If you find something I've worked on interesting, feel free to get in touch. Please contact me through LinkedIn if you do not already have my contact info. I look forward to working with you!