The problem
Video inputs differ in complexity, yet conventional 3D CNNs apply a largely fixed computation graph to every example. That wastes energy on resource-constrained edge devices.
What we did
We explored fine-grained, input-dependent pruning across frames, channels, and features. The related DANCE preprint describes activation variability amplification followed by a lightweight controller that adapts activation pruning to each input.
What we found
The approach reduces multiply-accumulate operations and memory accesses while preserving predictive performance. The DANCE preprint reports hardware validation on NVIDIA Jetson Nano and Qualcomm Snapdragon 8 Gen 1, demonstrating that adaptive pruning can translate into practical inference gains.
Paper
Mohamed Mejri, Ashiqur Rasul, Abhijit Chatterjee. “Dynamic 3D CNN Pruning: Joint Frame, Channel, and Feature Adaptation for Energy Efficiency on the Edge.” CVPR Workshop, 2025, as listed in the resume.