Francesco Tudisco – School of Mathematics Image Bayes Innovation Fellow: Francesco Tudisco What is your research focus? One of the main technological drivers of AI inefficiency, which my research addresses, is the use of unstructured parameter learning algorithms where full matrices or tensors need to be processed at each step, requiring huge computational power and storage overheads. The creation of more efficient low-parametric ML pipelines has the potential to dramatically lower associated costs, hardware, and energy requirements. The novel algorithmic technology I have developed combines techniques from high-performance scientific computing and model order-reduction of partial differential equations based on tensor networks, a powerful high- dimensional data decomposition model used in quantum systems that can provide exponential reduction in terms of number of parameters, while retaining guarantees of performance. Using this approach, we have achieved up to 95% model compression on large computer vision tasks, while achieving comparable or even better accuracy performance. What is your innovation idea? Through my research over the past two years, I have developed a novel algorithmic technology that reduces the number of parameters of large deep learning architectures, significantly reducing the cost and resources required to both train and deploy machine learning models. What problem will it solve? The problem I am addressing is the rapidly growing resource requirements of AI models. Machine learning (ML) advances have historically been driven by a combination of empirical trial and error and hardware advancements that, while allowing for the processing of ever-larger models, have resulted in expanding economic and social costs. Consequently, the present approach to AI development is no longer viable along its current trajectory. If mass-scale adoption of AI is to be sustainably and equitably actualized, it is critical that ML pipelines be made more efficient. What is the future of your research? My goal is to create a commercial platform that combines my research developments with other state-of-the-art model reduction techniques into production, and delivers services for customers. With this platform, large AI models can be deployed much more efficiently at considerably lower cost, delivering customers significant savings and expanding the number of enterprises capable of deploying customised AI models. This article was published on 2024-09-30