🔭 I work with deep learning and computer vision.
I also love systems programming especially the parts that relate to deep learning training infra. I mostly work with Python (PyTorch, JAX), love Rust and sometimes work with C/C++ and CUDA.
- motion transfer - Translate motion from videos to still images
I work on animation generation using deep learning. Dealing with stuff like motion transfer from video to 3D morphable models, lip animation from speech data and point cloud encoding using transformers (on a tight computational budget 😱).
Perviously I worked for an MLOps startup building an end2end MLOps and experiment management platform. There I worked on AutoML. Mainly, how to automatically optimize network architecture to fit training data and automating feedback loops for self-training object detection.
Before that I worked with the same people on optimizing video object detection inference for edge devices. Developing neural nets that can adjust their computation amount on the fly. The computation adjustment was done based on a gating mechanism that identified which parts of the frame can be reused. The other end of the system was adjusting the convolution computations to skip the reusable parts.
In Grad school I worked with Prof. Lior Wolf on adapting object detection neural net architectures to be more efficient on detecting text in historical documents