Embedded Files
Projects (Google Scholar)
My research spans various aspects of computing ranging from optimizing performance & energy-efficiency to reducing carbon emissions. In my dissertation, I studied various mobile and edge device workloads like audio/video streaming, fitness tracking, and ML apps. Developed and implemented optimizations targeting memory, compute, and network bottlenecks for resource constrained mobile devices equipped with ARM based processors and edge ML accelerators. I have also collaborated in research focussed on expanding in-memory computing paradigm to FPGA fabric in order to accelerate neural networks. Lastly, I have been working in carbon-emission modeling of emerging ML workloads to reduce the carbon footprint of their applications.
Projects (Google Scholar)
My research spans various aspects of computing ranging from optimizing performance & energy-efficiency to reducing carbon emissions. In my dissertation, I studied various mobile and edge device workloads like audio/video streaming, fitness tracking, and ML apps. Developed and implemented optimizations targeting memory, compute, and network bottlenecks for resource constrained mobile devices equipped with ARM based processors and edge ML accelerators. I have also collaborated in research focussed on expanding in-memory computing paradigm to FPGA fabric in order to accelerate neural networks. Lastly, I have been working in carbon-emission modeling of emerging ML workloads to reduce the carbon footprint of their applications.
Comprehensive Carbon Emission Modeling
Comprehensive Carbon Emission Modeling
Developed a comprehensive carbon emission model comprised of operational emission associated with the application, embodied emissions associated with the computing hardware, emissions from data transmission infrastructure, renewable energy sources, and the battery energy storage systems. Using this model, we proposed a framework that provides both an emissions predictor and a user-driven, emission-aware scheduler.
User-Driven Collaborative Recommendation Systems for Edge devices
User-Driven Collaborative Recommendation Systems for Edge devices
Developed a user-driven collaborative edge-datacenter recommendation system that intelligently splits the DNN recommendation models into cloud & edge to distribute the work and offload partial model computation to edge platforms. The edge model leverages local user information to reduce the memory/computational intensity by utilizing scratchpad, data reuse, and more. The proposed system offers user-data privacy while significantly improving applications' efficiency.
Hardware-friendly User-specific Machine learning for Edge Devices
Hardware-friendly User-specific Machine learning for Edge Devices
Proposed hardware-friendly technique to create small user-specific machine learning (ML) models on the edge device utilizing SIMD capabilities of CPUs and edge ML accelerator. User-specific ML models' reduced memory footprint and improved inference time significantly while ensuring user privacy.
End-to-end Machine learning-based Dynamic Sensing for IoT Devices
End-to-end Machine learning-based Dynamic Sensing for IoT Devices
Characterized IoT applications and developed a lightweight ML approach for multi-sensor edge devices to improve battery life based on user activity. The proposed approach established new communication between sensors where lightweight sensors guide compute at power-intensive sensors to reduce the overall computation at the device.
Interconnect Architecture for System-in-Package Based Low-Cost Edge Platforms
Interconnect Architecture for System-in-Package Based Low-Cost Edge Platforms
Studied data flow patterns in user-facing mobile applications to determine compute, memory, and network bottlenecks. Proposed an application cognizant 2.5D stacking-based System-in-Package (SiP) interconnect architecture and efficiently mapped data-flow patterns to leverage properties of 2.5D technology.
Page updated
Report abuse