Advanced Deep Learning Models (MERIT)

This course provides an introduction to advanced deep learning architectures and their applications in computer vision and generative artificial intelligence. It is part of the EU-backed MERIT project - an initiative aiming to boost the number of digital experts and reskill individuals, implemented in 6 European countries. MERIT is supported by the Digital Europe Programme of the European Union. 

About this course

Across three lectures, students explore key model families that have significantly shaped modern deep learning: object detection models, Generative Adversarial Networks (GANs), and Transformer-based architectures. 

• Lecture 1 – Object Detection: the first lecture focuses on object detection as a fundamental computer vision task. Students are introduced to the principles behind detecting and localising multiple objects within an image, including the concepts of bounding boxes, classification, localisation, and intersection over union (IoU). The lecture covers the evolution of modern detection architectures and the main differences between one-stage and two-stage approaches, together with practical considerations such as accuracy, computational cost, and real-time inference. 
• Lecture 2 – Generative Adversarial Networks (GANs): the second lecture introduces Generative Adversarial Networks and their application to generative modelling. Students learn about the interaction between the generator and discriminator networks and how adversarial training can be used to generate realistic synthetic data. Topics include the GAN training process, latent representations, image generation, common training challenges such as instability and mode collapse, and applications of GANs in areas such as image synthesis, data augmentation, and image-to-image translation. 
• Lecture 3 – Transformers: the third lecture introduces Transformer architectures and their growing importance beyond their original applications in natural language processing. Students explore the concepts of self-attention, multi-head attention, positional information, and the encoder-decoder architecture. The lecture examines how Transformers differ from traditional recurrent and convolutional architectures and introduces their applications in language, computer vision, multimodal learning, and generative AI. 

Learning Outcomes 

By the end of the course, students should be able to: 

• Explain the fundamental principles behind advanced deep learning architectures. 
• Describe the main components and training mechanisms of modern object detection models. 
• Understand the principles of adversarial training and explain how GANs can generate synthetic data.
• Identify the main challenges associated with training and evaluating GANs. 
• Explain the self-attention mechanism and the main components of Transformer architectures. 
• Compare different deep learning architectures in terms of their strengths, limitations, computational requirements, and typical applications. 
• Identify appropriate deep learning architectures for different computer vision and generative AI tasks.
• Interpret and critically discuss current applications of advanced deep learning models in research and industry. 
• Develop a broader understanding of how modern deep learning architectures are evolving towards increasingly general-purpose and multimodal AI systems.