Interests in ML training and inference performance improvements for real AI systems, with scalable resource-efficient architectures and sustainable optimizations (exploiting quantization / distillation / heuristics), using collected, curated, and existing multimodal data.
Research Works
NeurIPS’20:
EcoLight: Intersection Control in Developing Regions Under Extreme Budget and Network Constraints
Paper Code
NeurIPS’23:
AirDelhi: Fine-Grained Spatio-Temporal Particulate Matter Dataset From Delhi For ML based Modeling
Paper Code
ACM-JCSS’24:
FrugalLight: Symmetry-Aware Cyclic Heterogeneous Intersection Control using Deep RL with Model Compression, Distillation & Domain Knowledge
Paper Code
ACM-COMPASS’24:
WebLight: Deep RL based Multi-modal Intersection Control in Developing Countries without Reliable Cameras
Paper Code
ACM-COMPASS’22:
DynCNN: Application Dynamism and Ambient Temperature Aware Neural Network Scheduler in Edge Devices for Traffic Control
Paper
NeurIPS’23 (Workshop on Computational Sustainability)
RealLight: DRL based Intersection Control in Developing Countries without Traffic Simulators Paper