Skip to main content
Search

University of Leeds

3 related articles

Subscribe to updates
  1. arXiv

    Cutting three-quarters of visual tokens, APS-VLA keeps 95.75% robot-policy success and drops inference to 68.75 ms

    Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

  2. arXiv

    Skeleton action recognition gets a quantization plug-in: white-box attack success drops from near 100% to single digits, accuracy stays comparable

    Mainstream contrastive models for self-supervised skeleton-based action recognition had almost no defense against adversarial attacks, where tiny perturbations to skeleton data could sharply degrade recognition. QCLR adds quantization as a plug-in module at two levels of the encoding pipeline, mapping continuous inputs and features to discrete codebooks so small perturbations are absorbed and gradient propagation is broken; integrated into SkeletonCLR, AimCLR and ActCLR on NTU-60, NTU-120 and PKU-MMD, it cuts PGD and SMART attack success rates to single digits while keeping classification accuracy comparable to the original models.

  3. African exporters combining direct and indirect modes innovate more than single-mode exporters

    Using a Penrosian framework and World Bank Enterprise Surveys data on 4,254 African exporters surveyed between 2011 and 2020, the study tests whether pursuing direct and indirect exporting simultaneously yields stronger innovation than either mode alone, and finds that firms combining both modes innovate significantly more on product and process innovation, with the relationship robust across specifications and estimation techniques, while the four moderators tested only partly match theoretical predictions.

You’re all caught up