Research

My research focuses on developing sustainable and efficient machine learning systems, particularly for edge computing environments. I work at the intersection of computer architecture, machine learning, and sustainable computing to design energy-efficient systems for AI applications.

Current Research Areas

Sustainable Machine Learning Systems

I am developing novel techniques to reduce the energy consumption and carbon footprint of machine learning systems, particularly for edge devices. This includes:

  • Energy-aware neural network design and optimization
  • Sustainable training and inference methodologies
  • Carbon footprint measurement and reduction for ML workloads

Hardware-Software Co-design for Edge AI

My work explores the co-design of hardware and software to optimize machine learning performance on resource-constrained edge devices:

  • Custom accelerator architectures for neural network inference
  • Memory hierarchy optimizations for ML workloads
  • Hardware-aware neural network compression techniques

Efficient Neural Network Inference

I research methods to make neural network inference more efficient without sacrificing accuracy:

  • Model compression and quantization techniques
  • Sparse neural network architectures
  • Adaptive inference based on input complexity

Publications

For a complete list of my publications, please visit the Publications page.

Research Projects

NExUME - Adaptive Training and Inference for Intermittent Power

NExUME addresses the challenge of training and running deep neural networks under intermittent power conditions, typical in energy harvesting devices. The system provides adaptive mechanisms for both training and inference phases to maintain model performance despite power interruptions.

Usas - Sustainable Continuous Learning Framework

Usas is a comprehensive framework for sustainable edge computing that enables continuous learning on edge servers while optimizing energy consumption. The framework includes intelligent storage management, adaptive learning algorithms, and energy-aware resource allocation.

Origin - On-Device Intelligence for Activity Recognition

Origin enables on-device intelligence for human activity recognition using energy harvesting wireless sensor networks. The system combines efficient machine learning algorithms with energy harvesting techniques to provide continuous monitoring with minimal power consumption.

HoloAR - 3D Holographic Processing for Augmented Reality

HoloAR provides on-the-fly optimization of 3D holographic processing for augmented reality applications. The system optimizes computational resources and processing pipelines to enable real-time holographic rendering on resource-constrained devices.

Cocktail - Multidimensional Model Serving Optimization

Cocktail is a cloud-based system that provides multidimensional optimization for machine learning model serving. It optimizes across multiple dimensions including latency, throughput, accuracy, and resource utilization to provide efficient ML inference in cloud environments.

Kraken - Adaptive Container Provisioning

Kraken provides adaptive container provisioning for deploying dynamic DAGs in serverless platforms. The system automatically manages container resources and scaling to optimize performance and cost for complex computational workflows.

For more details on these and other projects, please visit the Projects page.