Skip to main content
Back

Georgia Tech's DANCE Cuts Video AI Power by 2.22x on Mobile Chips

A new framework from Georgia Tech dynamically prunes 3D convolutional networks per input, cutting energy use on edge devices by up to 1.47x with no accuracy loss.

Pro

Paper, Full Access
  • DANCE dynamically prunes 3D CNN frames, channels, and features per input for edge video inference (paper).
  • Two stages: activation variability amplification retrains the model, then a lightweight controller predicts pruning.
  • Delivers 1.37x speedup on Jetson Nano and 2.22x on Snapdragon 8 Gen 1 hardware.
  • Up to 1.47x higher energy efficiency versus prior state of the art with negligible accuracy loss.
  • Achieves 3.49x pruning rate on C3D with UCF101, beating previous best of 3.04x.
  • Code and data will be released after paper acceptance; not yet public.

Running video models on a phone or a small edge board is brutal on the battery. 3D convolutional networks used for action recognition, gesture classification, and video understanding scale their compute sharply when you move from images to video, and they burn the same energy on an easy clip as they do on a hard one. A group at Georgia Tech has proposed DANCE, which lets a 3D CNN decide, on the fly, how much of itself to actually run for each input.

Hardware validation on the NVIDIA Jetson Nano GPU and the Qualcomm Snapdragon 8 Gen 1 platform shows speedups of 1.37x and 2.22x respectively, with up to 1.47x higher energy efficiency compared to the prior state of the art. All of that with negligible or zero accuracy loss.

The bill that comes with 3D convolutions

2D CNNs are already expensive. Adding a time dimension for video makes things worse fast, since costs grow with network depth and scale up from 2D image processing to 3D video or point cloud processing. Structured pruning has been the standard answer, either trimming the model once or, more recently, adapting it to the input at inference time.

Granularity is where existing work falls short. Dynamic pruning has been applied mostly to 2D CNNs, where the focus is largely on coarse strategies such as channel pruning or layer removal. Those work, but they leave a lot of efficiency on the table. A boat drifting through calm water and the same boat crashing through rapids should not require the same amount of compute, yet most networks treat them identically.