Exploration of Accuracy, Performance and Resource usage trade-offs in Machine Learning Algorithms

  • Deployment of ML apps is challenging due to large number of computations at training and real-time latency require-ments & limited resources for inference.
  • Approximation can trade-off output quality for reduced execution time or same execution time with reduced resources i.e. can provide provide real-time latency for inference tasks and use less resources for intensive training tasks.
  • Employed approximations to explore this accuracy and performance & resource requirements trade-off.
  • Working on exploring possibilities to dynamically sacrifice accuracy in high load and limited resources scenarios.
Avatar
Neeraj Kulkarni
Ph.D candidate, Computer Systems Laboratory

My research interests include high-performance computer architecture, datacenters, system resource management.