Bayesian Prototypical Pruning for Efficient Video Understanding in Human-Robot Collaboration

Researchers at Northwestern Polytechnical University have proposed a novel end-to-end Bayesian framework, called Bayesian Prototypical Pruning (ProtoPrune), for efficiently pruning tokens in video understanding tasks. This method aims to improve the computation efficiency of video Transformers in human-robot collaborative (HRC) workspaces. By leveraging prototypical contrastive learning and variational dropout, ProtoPrune can achieve a pruning rate of 37.2% while retaining 92.9% of task performance, significantly improving computational efficiency.

Key Takeaways:

• Bayesian Prototypical Pruning (ProtoPrune) is a novel end-to-end Bayesian framework for token pruning in video understanding tasks.
• ProtoPrune leverages prototypical contrastive learning and variational dropout to improve robustness and efficiency.
• The method can achieve a pruning rate of 37.2% while retaining 92.9% of task performance using Uniformer and ActionCLIP.
• ProtoPrune offers a theoretically grounded and hardware-friendly solution for deploying video Transformers in real-world HRC environments.
• The research provides convergence analysis to ensure the stability of the method.
• The authors suggest that additional information can be obtained by contacting Bohua Peng, School of Automation, Northwestern Polytechnical University.

Statistics:

• 37.2% pruning rate achieved by ProtoPrune.
• 92.9% task performance retained by ProtoPrune using Uniformer and ActionCLIP.
• 13.9% improvement in computational efficiency achieved by ProtoPrune.

Sources:

• Peng, Bohua, et al. "Bayesian Prototypical Pruning for Transformers in Human-Robot Collaboration." Mathematics, vol. 13, no. 9, 2025, p. 1411, doi: 10.3390/math13091411. (Available at https://doi-org(sdpl.idm.oclc.org)/10.3390/math13091411)
• MDPI AG. Mathematics. (Publisher)