About Course

The seminar aims to deepen the understanding of the participants in current research problems at the intersection of deep learning and computer vision. The topics include, yet are not limited to: The evolution of CNN architectures (depth, layer composition, activation functions, drop out); Vision transformers (attention mechanism, basic and advanced architectures); Data augmentation techniques (basic techniques, MixUp, CutOut, AutoAug, RandAug, etc.); Semi-supervised learning (temporal ensembling, mean teacher, co-teaching, consistency loss and pseudo-labeling, MixMatch, ReMixMatch, etc.); Self-supervised learning (foundations of representation learning, pre-text tasks, contrastive loss, triplet loss, SimCLR, BYOL, SimSiam, etc.); Diffusion Models and Dall-E 2 (current breakthroughs in fusion of semantics of text and image, generative modeling); Explainable AI (clever hans phenomena, backprogation based method, gradient-based methods); Optimization theory of neural networks; Learning and generalization theory of neural networks (shortcut learning, forgetting/memorization events, layer importance, sample difficulty, learning paths, learning bias in texture/shape).

For the details about the course content, please visit the Moses page.

Contact Person: