IKT469 Deep Neural Networks (Spring 2027)

On successful completion of the course, the student should:

• Understand advanced concepts in deep learning, including why deep models generalize and how modern architectures are designed.
• Have knowledge of transformer models, multimodal representation learning, generative models, explainability methods, and advanced optimization strategies.
• Understand state-of-the-art architectures such as sequence models beyond attention and graph neural networks.
• Know how embedding works and how augmented systems combine neural networks with external knowledge sources.
• Design, implement, and evaluate advanced neural architectures across different data types and modalities.
• Build multimodal and retrieval-augmented systems using embeddings and external knowledge.
• Apply explainability methods to analyze and interpret deep learning models.
• Diagnose model performance and select suitable architectures for various tasks.
• Critically assess deep learning systems regarding performance, transparency, and computational trade-offs.

The course covers theoretical and practical aspects of neural networks:

• Advanced deep learning foundations and architectures: modern theory of deep networks, transformer models, contemporary neural architectures, and next-generation sequence and graph-based models.
• Representation learning and multimodality: embedding-based methods, multimodal alignment, and self-/semi-supervised learning.
• Generative and explainable AI: latent-variable and generative models, and practical explainability methods for neural networks.
• Knowledge-enhanced and optimized systems: retrieval-augmented generation, external knowledge integration, and advanced optimization and training strategies.

Combination of lectures, assignments, paper studies, lab, and report writing. The tasks are done individually or in small groups with group supervision.

The work load for the average student is approximately 200 hours.