ART-AI Seminar
We are pleased to have Alessandra Mileo, a tenured Assistant Professor in the Dublin City University (DCU) School of Computing, and a Funded Investigator at both the Insight & the I-Form Advanced Manufacturing Research Centres, join us for this ART-AI seminar, entitled ‘Knowledge Graph Analysis and Rule Learning meets Deep Representations for Explainability in Computer Vision’.
This seminar will take place online on Zoom on Tuesday 8th February 2022, 12.15pm-1.15pm (GMT).
Abstract
In this talk Alessandra will discuss her current agenda focused on using knowledge graphs analysis and rule learning to generate explainable representations of deep learning models in computer vision. She will provide an overview of her methodology in extracting knowledge from trained neural networks, validating such knowledge and injecting it back into untrained networks, with the goal of improving transparency and explainability. She will discuss how the approach can be used to detect, assess and correct bias and will raise questions and discussion prompts around how to decide what part of the model we can trust, and how to incrementally improve our explainable representation including domain experts in the loop.
Bio
Alessandra Mileo is a tenured Assistant Professor in the DCU School of Computing, and a Funded Investigator at both the Insight & the I-Form Advanced Manufacturing Research Centres. She has secured almost 1 million euros in funding including national (SFI, IRC), international (EU, NSF) and industry-funded projects, publishing 90+ papers, and she is an active PC member of over 20 conferences and journals. She is part of the National Centre for Research Training in AI and in ML at DCU.
Alessandra is particularly interested in neuro-symbolic computing as a way to design new comprehensive approaches to Explainable Artificial Intelligence. Specifically, she and her team are currently investigating a new approach to combining cognitive and neural learning and reasoning for holistic and human-centric Explainable AI for high-stake decision making, with particular interest in Diagnostic Imaging in clinical settings. Knowledge Representation, Rule Learning and Neural Networks can play a key role in designing machines that exhibit both cognitive and neural learning and reasoning, but many open questions remain on how to most effectively combine these two capabilities. Along with these challenges, such new approaches would need to bridge the gap between two faces of AI (connectionist and symbolic) that are historically diverging.
