While traditional feature engineering has been successful, modern machine learning increasingly relies on representation learning - automatically discovering informative features or representations from raw data. This seminar dives into advanced neural network-based approaches that learn dense vector representations capturing the underlying explanatory factors in complex, high-dimensional datasets.

The seminar will cover techniques like autoencoders, variational autoencoders, and self-supervised contrastive learning methods that leverage unlabeled data to learn rich representations. You'll learn about properties of effective learned representations like preserving locality, handling sparse inputs, and disentangling underlying factors. Case studies demonstrate how representation learning enables breakthrough performance on tasks like image recognition and natural language understanding. You'll gain insights into interpreting these learned representations as well as their potential and limitations.