I develop, train and optimize novel computer vision models for Semantic Segmentation, Pose Estimation, Object Detection, and Monocular Depth Estimation. I also spend time working on self-supervised training schemes for noise robustness and transfer learning.
I have extensive experience in HW-SW codesign, working closely with chip designers to enable new ML primitives on a unique hardware platform. This boils down to working up and down the complete software-hardware stack - that is: designing and training PyTorch models, running them through a custom dataflow compiler, and then running the compiled models on custom silicon. At each step, measuring the impact of architectural changes or perturbations made to the weights is the key to success.
Some other interesting work I’ve done:
- Created a library for efficiently training pytorch models and managing experiment configurations.
- Integrated the RayTune hyperparameter optimization suite with a custom hardware/software suite.
- Designed novel noise-conditioning SSL training schemes that selectively regularize noise-sensitive model layers.
- Created custom weight visualizations to qualitatively measure analog noise impacts on ML models.
In general, I work on optimizing and deploying novel computer vision models in challenging environments.
I also like kite-surfing, hiking (so does everyone else, I know), cooking, waterskiing, and eating.
This website is made with Hugo and the Congo theme.