Rajat Vadiraj Dwaraknath

CUDA and alpha

- Investigated the feasibility of various Mega Kernel approaches to simplify cuFFT planning and maintenance of kernel databases, as well as simultaneously increasing in-library kernel composability. - Prototyped an FFT-specific "Kernel LEGO" approach for dynamic kernel generation using Just-In-Time Link Time Optimization (JIT LTO). - Resulted in a 6.8x reduction in binary size of a database component while facilitating the explosive growth of supported FFT functionalities. - Developed a C++ and Python meta-programming framework for fast prototyping and testing of kernel generation strategies.

- Developed a faster Neural Message Passing framework for offline video tracking by modifying the algorithm to operate at the tracklet level. - Studied and extended the Snorkel semi-supervised learning framework to the object detection problem. - Deployed a highlight extraction and motion detection model to production.

- Worked on Predictive Modeling of partial differential equations using machine learning and automatic differentiation. - Analyzed the effect of gradient descent on different approximation methods in this context, and identified an implicit regularization effect due to early stopping that allowed ill-conditioned models to perform comparably to neural networks. - Involved performing large scale benchmarks in parallel on the Sherlock Computing Cluster.

- Worked on Anomaly detection in high frequency time series data obtained from Industrial IoT sensors. - Implemented Statistical and Machine Learning algorithms like LSTMs and integrated them into an existing data pipeline.