Department of Computing students receive deep learning awards

Students on the Department of Computing's Deep Learning course gained experience in its theoretical foundations and cutting-edge applications.

In this year’s fourth-year Deep Learning course, approximately 400 students led by Dr Yingzhen Li and Professor Bernhard Kainz, with support from Dr Harry Coppock of the UK AI Safety Institute, were guided through both the theoretical foundations and cutting-edge applications of deep learning. Spanning fundamental principles, architectures, and optimization strategies, the course culminates in a unique generative AI project inspired by HBO’s Silicon Valley TV series.

In this playful yet rigorous assignment, students initially develop a classifier capable of distinguishing "hot dog" images from "not hot dog" images, gaining hands-on experience with data preprocessing, model selection, and generalization challenges. Building on this foundation, they then design generative models to produce entirely new hot dog images. Despite the humour, this project introduces complex topics such as dataset bias, generative AI models, and the inherent constraints of synthetic data. Working on modest single-GPU compute setups, far removed from the massive computational resources of industry giants, students perfect their creativity, problem-solving skills, and practical understanding of how deep learning models learn and generalise.

This year’s competition for the most compelling hot dog images highlighted exceptionally resource-efficient diffusion models, a type of AI models that are capable of synthesising new images from pure noise. Similar models have gained significant recognition through recent advancements such as StableDiffusion and DALL·E, the latter now integrated into OpenAI’s ChatGPT.

Congratulations to our winners:

Winner: James Myles

Runners-up: Ryan Patel and Zhiming Hong