Dr Nitisha Jain is a Postdoctoral candidate in Computer Science in the Department of Informatics at King's College London. She pursued her PhD on Knowledge Graphs from the Hasso Plattner Institute (University of Potsdam). Prior to this, she worked at the Max Planck Institute of Informatics (Saarland University) and was a Research Scientist at IBM AI, India research lab. Previously, she completed Masters in Research from Indian Institute of Science (IISc) in 2015 with a focus on Cloud Computing.

Her primary research interests are related to Knowledge Graphs, Semantic Web and Machine Learning and recent publications can be found at https://nitishajain.github.io/. She is actively involved in the EU Horizon MuseIT project (https://www.muse-it.eu/).

Abstract:
Knowledge graph embeddings that generate vector space representations of knowledge graph triples, have gained considerable popularity in past years. Relying on the presumed semantic capabilities of the learned embeddings, they have been leveraged for various other tasks such as entity typing, rule mining and conceptual clustering. However, a critical analysis and evaluation of these embeddings in terms of semantic representation of the underlying entities and relations reveals their limitations. The semantic representation in the knowledge graph embeddings is not universal but restricted to a small subset of the entities based on dataset characteristics. This can be improved by way of incorporating ontological reasoning during the training of the embeddings. The ReasonKGE approach identifies incorrect predictions produced by a given embedding model dynamically via symbolic reasoning and feeds them as negative samples for retraining the model. This method demonstrates the gains in accuracy and semantic consistency of facts produced by embedding models, thus improving their overall semantic capabilities. The fact predictions from embeddings and the semantic features encoded in such models play an important role in multi-modal knowledge completion tasks. Recent works on finding semantic interpretations in embedding techniques hold promise and could pave the way for explainable multi-modal predictions from knowledge graph embeddings. 

Dr Fabio Pizzati is a postdoctoral researcher in the Torr Vision Group at the University of Oxford, working with Prof. Phil Torr. He is currently working on large scale diffusion models for image generation, and continual learning.

Generative models, such as GANs and diffusion models, have yet to fully incorporate the nuances of physical information. In this presentation, we address this oversight from two distinct yet opposite angles. First, we’ll showcase how incorporating physical models into GANs can enhance their control and output quality. Conversely, we’ll explore how advanced large scale diffusion models, especially Stable Diffusion, can be harnessed to discern physical attributes from natural images.

Title: Innovating in Medical Device Development



Abstract:

Development of Software as a Medical Device have shown progress in the last two decades due to rapid innovation and adaption of new technologies, good documentation and educational resources for students, researchers, academics, engineers and clinicians. However, research-driven technologies bring other challenges when using the latest generation of hardware and software, including the validation of new algorithms and the standardisation of data quality and data privacy to mention but a few. These challenges raise the question of how to quickly adopt the latest technologies (e.g., artificial intelligence, augmented reality, high-performance computing, etc) while still complying with relevant quality standards. Hence, Miguel is presenting his work and challenges on finding the right balance between innovation and regulations on clinical translation of different case studies in Ultrasound Imaging: needle tracking, echordiography classification in low-resource countries, GAN-based and Diffusion-based Fetal Ultrasound Image synthesis and Open-Source Software technologies aligning with medical regulations.



Bio:

Miguel is a Research Engineer at University College London within the Advanced Research Computing Centre and WEISS where he is advancing AI-based Surgical Navigation tools. Previously, he was a Research Associate at King’s College London where he advanced research in Ultrasound-Guidance Procedures and AI-enabled echocardiography pipelines. In 2019, he was awarded a Ph.D. degree in Computer Engineering from the University of Birmingham, researching “Nonlinear Analysis to Quantify Movement Variability in Human-Humanoid Interaction”. His primary research interests are in developing data-centric AI algorithms for Medical Imaging, MedTech, SurgTech, Biomechanics and clinical translation. Additionally, his work includes generative models for fetal imaging, sensor fusion data from time-series and medical imaging, real-time AI for echocardiography, image-guided procedures, AI-based surgical navigation tools, and child-robot interaction in low-resource countries.



* This is a hybrid event. You can also join remotely. Please see Teams link on the OxfordXML webpage (see URL provided in Further Information).

Deep learning models, especially those large- scale and high-performance ones, can be very costly to train, demanding a considerable amount of data and computational resources. As a result, deep learning models have become one of the most valuable assets in modern artificial intelligence. For instance, it can cost $1.6 million to train a BERT model on Wikipedia and Book corpora. It is thus of utmost importance to protect deep learning models from unauthorized duplication or reproduction. This talk will be about threats to deep learning copyright and corresponding defence techniques on copyright protection of deep learning models.