Architecture and GAN-based training procedure for obtaining a fast neural network which enable better bitrate performances respect to the H.265 codec for the same quality, or better quality at the same bitrate. read more of Fast SR-UNet

d-Simplex classifiers achieve stationary, compatible features, enabling seamless model updates in retrieval systems. read more of Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements

A novel dataset of dynamic sequences of 3D face models, where a combination of synthetic and real identities exhibit an unprecedented variety of 4D facial expressions, with variations that include the classical neutral-apex transition, but generalize to e... read more of CL2R: Compatible Lifelong Learning Representations

Using stationary representations, CoReS trains models to obtain compatible representations, eliminating costly re-indexing in retrieval systems during upgrades. read more of CoReS: Compatible Representations via Stationarity

Mitigating forgetting in continual representation learning using contrastive supervised distillation. read more of Contrastive Supervised Distillation for Continual Representation Learning

Using fixed classifiers derived from regular polytopes to enhance neural network efficiency and accuracy by generating stationary, maximally-separated feature representations. read more of Regular Polytope Networks

Elastic Feature Consolidation (EFC), exploits a tractable second-order approximation of feature drift based on an Empirical Feature Matrix (EFM). read more of Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning

We propose a more realistic, physics-based color data augmentation - which we call Planckian Jitter. read more of Planckian jitter: enhancing the color quality of self-supervised visual representations

We propose a two-stage learning baseline with a learnable weight scaling layer for reducing the bias caused by long-tailed distribution in LT-CIL and which in turn also improves the performance of conventional CIL due to the limited exemplars. read more of Long-Tailed Class Incremental Learning

Neural network based on an end-to-end trainable working memory, which acts as an external storage where information about each agent can be continuously written, updated and recalled read more of Smemo: Social memory for trajectory forecasting