Checkpointing in Practice for Memory-Efficient Training on the Edge

Last updated on Aug 15, 2022
IEEE Website Slides
  • The research involved developing a novel and systematic method to enable the training of Deep Neural Networks such as ResNet, MobileNet, VGG, DenseNet, etc. on memory-constrained Edge devices.
  • Performed experiments with training Deep Neural Networks in devices with 1GB RAM and was successfully able to run such Architectures in a Raspberry Pi Board, which was otherwise not possible.
  • Optimized training of MobileNet and ResNet-18 architectures and thereby reduced memory consumption by a factor of 2.6 and 1.8 respectively without loss in predictive power.
DL ML Systems
Naveen Elango
Naveen Elango
Product Manager, Paytm

My research interests include distributed robotics, mobile computing and programmable matter.