Sheffield Machine Learning Seminar Series 2016/17 [Upcoming | Past]
Date: 21 June 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: From Random Projections to Learning Theory and Back;
Speaker: Dr Ata Kaban, University of Birmingham.
We consider two problems in statistical machine learning -- an old and a new:
(1) Given a machine learning task, what kinds of data distributions make it easier or harder? For instance, it is known that large margin makes classification tasks easier.
(2) Given a high dimensional learning task, when can we solve it from a few random projections of the data with good-enough approximation? This is the compressed learning problem.
This talk will present results and work in progress that highlight parallels between these two problems. The implication is that random projection -- a simple and effective dimensionality reduction method with origins in theoretical computer science -- is not just a timely subject for efficient learning from large high dimensional data sets, but it can also help us make a previously elusive fundamental problem more approachable. On the flip side, the parallel allows us to broaden the guarantees that hold for compressed learning beyond of those initially inherited from compressed sensing.
Ata Kaban is a senior lecturer in Computer Science at the University of Birmingham UK, and EPSRC Early Career Fellow. Her research interests include statistical machine learning and data mining in high dimensional data spaces, algorithmic learning theory, probabilistic modelling of data, and black-box optimisation. She authored / co-authored 80 peer-reviewed papers, including best paper awards at GECCO'13, ACML'13, ICPR'10, and a runner-up at CEC'15. She was recipient of an MRC Discipline Hopping award in 2008/09. She holds a PhD in Computer Science (2001) and a PhD in Musicology (1999). She is member of the IEEE CIS Technical Committee on Data Mining and Big Data Analytics, and vice-chair of the IEEE CIS Task Force on High Dimensional Data Mining.
Date: 27 September 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Speaker: Professor Winston Hide, University of Sheffield.
Date: 29 March 2017 (Wednesday); Time: 10:00am-11:00am; Venue: Ada Lovelace (Regent Court COM-108);
Title: Stochastic (Partial) Differential Equations and Gaussian Processes;
Speaker: Prof. Simo Särkkä, Aalto University, Finland.
Stochastic partial differential equations and stochastic differential equations can be seen as alternatives to kernels in representation of Gaussian processes in machine learning and inverse problems. Linear operator equations correspond to spatial kernels, and temporal kernels are equivalent to linear It\^o stochastic differential equations. The differential equation representations allow for the use of differential equation numerical methods on Gaussian processes. For example, finite-differences, finite elements, basis function methods, and Galerkin methods can be used. In temporal and spatio-temporal case we can use linear-time Kalman filter and smoother approaches.
Prof. Simo Särkkä received his Master of Science (Tech.) degree (with distinction) in engineering physics and mathematics, and Doctor of Science (Tech.) degree (with distinction) in electrical and communications engineering from Helsinki University of Technology, Espoo, Finland, in 2000 and 2006, respectively. From 2000 to 2010 he worked with Nokia Ltd., Indagon Ltd., and Nalco Company in various industrial research projects related to telecommunications, positioning systems, and industrial process control. From 2010 to 2013 he worked as a Senior Researcher with the Department of Biomedical Engineering and Computational Science (BECS) at Aalto University, Finland.
Currently, Dr. Särkkä is an Associate Professor and Academy Research Fellow with Aalto University, Technical Advisor and Director of IndoorAtlas Ltd., and an Adjunct Professor with Lappeenranta University of Technology. In 2013 he was a Visiting Professor with the Department of Statistics of Oxford University and in 2011 he was a Visiting Scholar with the Department of Engineering at the University of Cambridge, UK. His research interests are in multi-sensor data processing systems with applications in location sensing, health technology, machine learning, inverse problems, and brain imaging. He has authored or coauthored ~80 peer-reviewed scientific articles and has 3 granted patents. His first book "Bayesian Filtering and Smoothing" and its Chinese translation was recently published via the Cambridge University Press. He is a Senior Member of IEEE and serving as an Associate Editor of IEEE Signal Processing Letters from August 2015.
Date: 5 April 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: On the (statistical) detection of adversarial examples;
Speaker: Kathrin Grosse, Saarland University, Germany.
Imagine meeting a dear friend and thinking he is your mother, because he is wearing some glasses. Wait, what? Out of question for most of us, however reality for many machine learning models. So called adversarial examples are original samples where an adversary computes an optimal perturbation, leading to a different classification. To humans, however, the two examples are in most cases not distinguishable. The automated detection of such adversarial examples remains an open problem, since the perturbations are a consequence of an inherent property of all classifiers: the gradient of the decision function.
In this talk, we will first briefly review how adversarial examples are computed (using the example of malware data). We then move to our work on how to detect adversarial examples, presenting two approaches. One confidently detects adversarial examples, however only when presented in a batch. Another approach works as well for a single example, yet is does not work as a reliable defence in all cases: In some cases, it only increases the cost of the attack.
I studied cognitive sciences at Osnabrück (Lower Saxony, Germany). I specialized in computer science, AI and neurobiology. In my term abroad at the Universidad del Sur in Bahia Blanca, Argentina, we (joint work with Carlos Chesñevar) started to work on opinion mining on Twitter. Around then I decided to work in Data Mining and ML, and continued my studies at Saarland University (since there is a lot of Data Mining/ML specialists there). I wrote my master thesis on text mining in Jilles Vreeken's group, who does Minimum Description Length based exploratory data analysis. Although always enjoying data mining and machine learning in itself, I became interested in security, particularly the security of ML. Due to this interest, I started working (as a phd student) on this topic at Michael Backes group at CISPA.
Date: 26 April 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: Computational challenges in genomics and personalized medicine;
Speaker: Dr Dennis Wang, University of Sheffield.
Recent technological advances in the high-throughput profiling of DNA, RNA and proteins in human tissue have lead to a wealth of genomic data and better understanding of complex diseases, such as cancer, diabetes and neuro-degeneration. This, however, has also lead to two major computational challenges of classification and feature selection. In this talk, I will highlight examples where supervised and unsupervised clustering approaches have been applied by drug developers aiming to produce "personalized" medicines from genomic data. I will also point out further machine learning problems and areas for collaboration with Sheffield's medical and biological research communities who are dealing with increasing amounts of genomic data.
I graduated from the University of British Columbia (Vancouver, Canada) and completed my undergrad dissertation at the European Bioinformatics Institute (Cambridge, UK). I then moved to the University of Cambridge for an MPhil working on boolean networks with Dr. Jasmin Fisher and Dr. Andrew Phillips (Microsoft Research), and a PhD working on statistical genetics with Prof Lorenz Wernisch and Prof Willem Ouwehand (MRC Biostatistics Unit).
Following the completion of my PhD in 2012, I undertook postdoctoral training to build a genomics core and identify biomarkers at the Princess Margaret Cancer Centre (Toronto, Canada) with Prof Ming-Sound Tsao and Prof Frances Shepherd. I was promoted to a staff scientist in 2013 to coordinate the genomic profiling and bioinformatics analysis of patient tumors. With a greater interest in drug development, I went back to Cambridge in 2014 to join the early drug discovery division of AstraZeneca where I developed computational methods to identify pharmacological biomarkers that predict drug response. I joined the University of Sheffield in 2016 as a lecturer to further establish genomics and bioinformatics as cornerstones within the education and research programmes at the medical school.
Date: 03 May 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: RSS-based Indoor Localization using Gaussian Processes;
Speaker: Dr Roland Hostettler, Aalto University, Finland.
Location-based services such as augmented reality benefit greatly from accurate position estimates. In outdoor environments, global navigation satellite systems often offer adequate performance for a broad range of applications. However, performance of these systems degrades quickly when moving to the indoors due to the complex building structures affecting the satellite signals. Additionally, in indoor environments, an error of only a few meters can be the difference between two rooms or even floors. In this talk, we will discuss an alternative localization approach based on fingerprint maps of radio signals. The method uses Gaussian processes to model the radio landscape and can be used with, for example, WiFi or Bluetooth signals.
Roland Hostettler received the Dipl. Ing. degree in electrical and communication engineering from Bern University of Applied Sciences, Switzerland in 2007, and the M.Sc. degree in electrical engineering and Ph.D. in automatic control from Luleå University of Technology, Sweden in 2009 and 2014, respectively. From October 2014 to January 2016, he was a research associate at the Control Engineering Group at Luleå University of Technology. Since February 2016 he is a postdoctoral researcher at the Department of Electrical Engineering and Automation at Aalto University, Finland.
His main research interests are statistical signal processing in general and parameter inference, state estimation, and sensor fusion in particular with applications in localization, tracking, as well as activity and health monitoring.
Date: 10 May 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Pool Seminar Room G03, 9 Mappin Street;
Title: Some frontiers in deep reinforcement learning;
Speaker: Dr Tom Schaul, Google DeepMind.
Two desiderata for general intelligence are performance and generality. The first requires acting to achieve goals or solve problems; the second asks for agents that are competent on a diversity of tasks, or at least can learn to become so -- with minimal teaching signal, if possible. After decades of focus on performance, recent research has started to emphasize the aspect of generality, building on two key ingredients, namely reinforcement learning (RL) and deep neural networks. I will discuss some of their strengths and weaknesses, put recent breakthroughs into context (e.g. AlphaGo, DQN), and sketch outs some ongoing directions that could push generality even further.
Tom Schaul is a senior researcher in reinforcement learning at DeepMind. He did his PhD with Jürgen Schmidhuber at IDSIA and his Postdoc with Yann LeCun at NYU. He has published many areas of AI, including deep learning, optimization algorithms, artificial curiosity, evolutionary algorithms, and most recently on deep and hierarchical RL. He thinks that substantial progress on general AI is possible, and that games are perfect benchmark domains for that.
Date: 07 June 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: Statistical long-term excitatory and inhibitory synaptic plasticity;
Speaker: Dr Tim Vogels, University of Oxford.
Long-term modifications in neuronal connections are critical for reliable memory storage in the brain. However, pre- and postsynaptic components can make synapses highly unreliable. How synaptic plasticity modifies this variability is poorly understood. Here we introduce a theoretical framework in which long-term plasticity performs an optimisation of the postsynaptic response statistics constrained by physiological bounds. In this framework of statistical long-term synaptic plasticity the state of the synapse at the time of plasticity induction determines the ratio of pre- and postsynaptic changes. When applied to plasticity of excitatory synapses, our theory explains the observed diversity in expression loci of individual hippocampal and neocortical potentiation and depression experiments. Moreover, our theory predicts changes at inhibitory synapses that are bounded by the mean excitation, which suggests an efficient excitation-inhibition balance in the brain. Our results propose a principled view of the diversity in expression loci of long-term synaptic plasticity observed in a wide range of slice experiments and reveal a statistically optimal, excitation-inhibition balance in the intact brain.
Tim Vogels studied physics at Technische Universität Berlin and neuroscience at Brandeis University as a Fulbright Scholar. He received his PhD in 2007 in the laboratory of Larry Abbott. After a postdoctoral stay as a Patterson Brain Trust Fellow with Rafa Yuste at Columbia University, he became a Marie Curie Reintegration Fellow in the laboratory of Wulfram Gerstner at the École Polytechnique Fédérale de Lausanne (EPFL). Tim was awarded the Bernstein Award for Computational Neuroscience in 2012.
Tim Vogels arrived at Oxford in 2013 and is establishing a research group in theoretical and computational neuroscience within the Centre of Neural Circuits and Behaviour. As a computational neuroscientist, he builds conceptual models to understand the fundamentals of neural systems at the cellular level. His research group is funded by a Sir Henry Dale Fellowship of the Wellcome Trust and the Royal Society and part of the neurotheory initiative at the University of Oxford.
Date: 14 June 2017 (Wednesday); Time: 3:00pm-4:00pm; Venue: Ada Lovelace (Regent Court COM-108);
Title: Factorised Gaussian Process Models;
Speaker: Dr Carl Henrik Ek, University of Bristol.
Regression is the task of relating an input variate to an output domain by the means of a function. To learn the mapping from data means that we are faced with the daunting task of specifying a distribution over the space of functions. Gaussian process priors allow us to do just this in an interpretable and flexible manner. However, for many types of data the relationship cannot be described by a function as there are multiple parts of the output domain corresponding to the same input location. In this scenario we often have to resort to latent variable models in order to capture the relationship which are often characterised by expensive and challenging inference scenarios.
In this talk I will describe a set of different approaches to modelling in the above described scenario. Our idea is that we can build models that learns a factorisation of the variations in the data such that we can simplify the inference problem. I will exemplify some of these models on real-data from robotics and computer vision.
Dr. Carl Henrik Ek is a lecturer at the University of Bristol. His reasearch focuses on developing computational models that allows machines to learn from data. In specific he is interested in Bayesian non-parametric models which allows for principled quantification of uncertainty, easy interpretability and adaptable complexity. He has worked extensively on models based on Gaussian process priors with applications in robotics and computer vision.