From model failure to human-in-the-loop computer vision: James Burgess @ Bayezian

​We’re back.

​Computer vision projects are often approached by asking which model is best suited to the problem. In practice, that can be one of the less important decisions. This talk follows the development of a computer vision system for analysing beating myocardiocytes, where an initially promising modelling approach struggled once it met the realities of limited training data, difficult image annotation and a commercially constrained project timeline. Rather than continuing to optimise the same approach, the project shifted towards an iterative human-in-the-loop workflow using Cellpose, allowing more efficient labelling and model training to improve together.

​These lessons were then used to develop our own human-in-the-loop labelling and training approach, designed to make expert annotation and model improvement part of the same workflow. The talk will also explore how these principles carry over into the growing use of foundation models and LLM vision models, and why increasingly capable models do not remove the need for careful data, workflow and project design.

​James is an AI Research Consultant at Bayezian. He has an MSc in Artificial Intelligence and MSci in Neuroscience from Sussex.

​The usual gossip session will happen afterwards at the pub.