<ul><li>Biological brains learn continually from unlabeled data and integrate specialized information from sparsely labeled examples without compromising generalization.</li><li>Machine learning methods face catastrophic forgetting in natural learning settings, where supervised specialist fine-tuning degrades performance on original tasks.</li><li>Task-modulated contrastive learning (TMCL) is introduced, inspired by neocortical biophysical machinery, integrating top-down information continually and unsupervisedly using predictive coding principles.</li><li>Experiments demonstrate improvements in class-incremental and transfer learning over state-of-the-art unsupervised methods, suggesting the importance of top-down modulations in balancing stability and plasticity.</li></ul>

Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning

Discover more