AI talks by a group at Univ. College London
July 20 (Fri.)
National Institute of Informatics
20F Meeting Room, NII
Evolving bio-inspired AI
Spiking neural networks, thanks to their sensitivity to the timing of the inputs, are a promising tool for unsupervised processing of spatio-temporal data. However, they do not perform as well as the traditional machine learning approaches and their real-world applications are still limited.
Various supervised and reinforcement learning methods for optimising spiking neural networks have been proposed, but more recently the evolutionary approach regained attention as a tool for training neural networks. Here, we describe a simple evolutionary approach for optimising spiking neural networks. This is the first published use of evolutionary algorithm to develop hyperparameters for fully unsupervised spike timing- dependent learning for pattern clustering using spiking neural networks. Our results show that combining evolution and unsupervised learning leads to faster convergence on the optimal solutions, better stability of fit solutions and higher fitness of the whole population than using each approach separately.
Katarzyna Kozdon is an AI researcher with a strong interdisciplinary background. She is currently pursuing a PhD in Spiking Neural Networks, and works as a teaching assistant for Maths, Informatics and Computational Biology (SysMIC) and Robotics Programming courses at University College London. She holds a Master degree in Neuroscience, and was a doctoral advisor to the gold-winning UCL international Genetically Engineered Machine competition team (iGEM 2016).
Agent-Based Modelling of Colorectal Cancer
Intestinal glands in the small intestine and colon, or intestine crypts, are an important example of tissue homeostasis regulated by the extracellular environment. The crypts are invaginated structures made of a layer of cells that help absorb nutrients from passing food. However, they are continuously worn away by this process and are being continually renovated by Stem Cells at the bottom of the crypt. These Stem Cells divide to replace worn cells and may even displace other stem cells so that at a given time the whole crypt becomes monoclonal- a descendant of one single Stem Cell. Colorectal Cancer, the second leading cause of cancer-related death in Europe and North America, is thought to start with a mutation of one Stem Cell at the base of the intestinal crypt; which then expands within the crypt until the crypt is composed of monoclonal cells. The time to monoclonality therefore offers a key metric for the successful establishment of mutations.; however, the biggest biological contributor to this feature is highly debated. Three key hypotheses have been put forwards, which we investigated with ALife methods.
We have abstracted key biological features and modelled them in a bottom-up Agent-Based Model that allowed us to study the biological first principles that rule the fixation of mutations, offering key spatial and temporal understanding of this process. Our results show that the number of basal Stem Cells have a direct influence on the fixations of mutations and suggesting a lesser role for extracellular influences, while proposing the existence of a threshold to the contribution of cell side displacement.
Dr Arturo Araujo is a Senior Research Fellow at Braintree's Research Department. There, he leads a team which examines systems related to natural life, its processes, and its evolution, through the use of simulations with computer models, robotics, and biochemistry. Arturo is a physicist trained into a modeller of biological complexity at University College London's CoMPLEX program. He focused his PhD on modelling aneuploidy -chromosomal aberrations- in cancer; a key feature that is difficult to assess in vivo and in vitro. Following the successful completion of this research, he was recruited by the Integrated Mathematical Oncology (IMO) department at H. Lee Moffitt Cancer Center, where he used all of his interdisciplinary training and viewpoints to successfully bridge the gap between theory, experiments and the clinic with integrated computational models in tight collaboration with biologists and clinicians. In his research at Braintree, he develops cutting-edge computational techniques to understand cancer dynamics and design better treatments at the interphase between AI, biology and the clinic. In his spare time, Arturo likes to do art.
A Platform for Graph Analytics-based Modular Machine Learning and Artificial Intelligence (GAMMA)
An overview of the GAMMA platform, its motivations and design. This talk will provide a summary of problems that drive research in graph-based machine learning with discussion of some of the latest algorithms and advances, while also giving consideration to how such algorithms can be computed and scaled for learning over massive datasets.
Peter Meltzer is a Research Associate at Braintree Ltd. responsible for leading the research and development of the GAMMA platform. He is also studying a PhD in machine learning on graph-structured data at UCL, where his research interests include scalable machine learning, graph-based and distributed computation, and knowledge representation.
Agent-based Modelling of Human Personality in Collaboration
Collaboration is an essential aspect of human interaction. Despite being mutually beneficial to everyone involved, it often fails due to behaviour differences as individuals process information, form opinions, and interact with each other. As such, in order to understand collaboration, it is necessary to consider the psychology of the individuals involved. We propose an agent-based model of collaboration that incorporates human personality.
In the model, the shared goal is abstracted as a shared optimisation task, and personality differences are abstracted as strategies for moving within, interpreting and sharing information about the solution space. To test the effectiveness of this approach, we investigated performance of groups of individuals with contrasting personality types, as they collaboratively solve a problem with varying degrees of noise. The model predicted significant differences between personality types and the predictions were corroborated by the literature.
Dr Soo Ling Lim is Chief R&D Officer at Braintree Ltd, Honorary Research Associate at the University College London Department of Computer Science, Visiting Assistant Professor at the National Institute of Informatics, and Fellow of the Higher Education Academy in UK. Her research focuses on understanding human collaboration relating to software development. It encompasses stakeholder analysis, agent-based modelling, personality psychology, requirements elicitation, crowdsourcing, and social network analysis. Soo Ling received a PhD in large-scale software requirements engineering from the University of New South Wales in Sydney, Australia in 2011, and a Bachelor of Software Engineering from the Australian National University in 2005.
Fuyuki Ishikawa ( f-ishikawa [at] nii.ac.jp )