cuDNN heuristics: improving the performance of the cuDNN library.

Approximate video analytics on mobiles• Designed novel approximate DNN models and object detection frameworks.• Designed novel schedulers for the approximate system to be content and resource contention aware.• Designed efficient systems with low overhead.• First-author papers on premier conferences and journals -- SenSys, TOSN, and ISM.Workload profiling and scheduling in the data-center• Built tools to measure the memory contention between multiple collocated workloads in the data-center.• Studied the memory contention between workloads and built prediction models for incoming workloads.• Designed the scheduling algorithm by putting multiple workloads on one host and guaranteeing the QOS of them.• First-author paper at 2018 ACM/IFIP International Middleware Conference.

Graph Embedding Learning on Search Data• Developed the training and evaluation pipelines of node2vec models to learn the entity embeddings from the Search data.• Improved the recall of predicting the target groups from 31.9% to 40.3% with the trained models.

Improving creators' experience on content creation• Developed closed-loop feedback on video/image ad to the creators based on the performance data.• Developed a general framework for any creation and evaluation apps to reduce the cost of searching an optimal variant.

Utility pole detection with drone cameras• Built image classification system for detecting utility poles using Single Shot MultiBox Detector (SSD) model.• Developed software simulation to verify the robustness of image classification models.• Sped up the inference of the deep neural networks (DNN) with the edge clouds.

Erasure-code repair performance in the distributed object storage system (ceph) • Designed and implemented a novel repair protocol to enable parallel repair of the data chunks.• Identified the performance bottleneck and reduced the repair latency.

Topology analysis in vehicular networks• Analyzed the temporal link topology of urban large-scale vehicular networks.• Proposed a novel opportunistic reachability graph to characterize the topology and an efficient algorithm to compute.Coverage analysis in cellular networks• Analyzed the spatial and temporal coverage pattern of urban large-scale cellular networks.• Analyzed the signal strength especially inside buildings.

RF sensing for human dehydration• Designed a novel method to sense the moisture content with 60 GHz, 2.4 GHz radios.• Processed the noisy data to extract moisture content – sensitivity reaching one drop (0.02 milliliter).