This course covered the theory and practice of modern neural networks: why deep models generalize, how contemporary architectures are designed, and how to build systems that combine neural networks with external knowledge.
Transformers and Modern Architectures
We studied transformer models in depth alongside next-generation sequence models beyond attention and graph neural networks. The focus was architectural judgment: diagnosing model performance and selecting suitable architectures for a task, rather than defaulting to whatever is newest.
Representation Learning and Multimodality
The representation module covered embedding-based methods, multimodal alignment, and self- and semi-supervised learning. Building on my earlier multimodal genre-classification work, I learned how modern embedding spaces make cross-modal systems composable.
Generative and Explainable AI
We worked with latent-variable and generative models, and applied practical explainability methods to analyze and interpret deep networks. Critically assessing systems on performance, transparency, and computational trade-offs was a stated learning outcome, and a habit I've carried into my document-security work, where explainability is a product requirement, not a nice-to-have.
Knowledge-Enhanced Systems (RAG)
The final module covered retrieval-augmented generation and external knowledge integration, plus advanced optimization and training strategies for making these systems efficient.
Course Projects
Two public projects came out of this course: Deeper-Neural-Networks, implementing and evaluating advanced architectures across data types and modalities, and uia-rag-chatbot, a retrieval-augmented chatbot that grounds a language model in university course material using embeddings and external knowledge sources.