MPhil-PhD transfer presentation

When: Thu, 30 Nov 2017, 12noon
Where: EM01

Who: Simon Odense, City, University of London

Title: Compact Rule Extraction from Probabilistic Neural Networks

Abstract: I will discuss new techniques for extracting M-of-N rules from restricted Boltzmann machines. I will begin by discussing rule extraction and its importance for explainable AI before describing a method for extending the extraction of conjunctive confidence rules to the more compact M-of-N rules. Comparative results between extraction algorithms will be presented and several possible experimental applications will be discussed. I will also discuss methods of assigning confidence values to extracted rules in a logical way and how to factor in compactness/interpretability in the extraction process.

All welcome!

Machine Learning seminar

When: Fri, 20 Oct 2017, 1pm
Where: AG24a, College Building

Who: Alessandro Abate, University of Oxford

Title: Data-driven and model-based quantitative verification of physical systems

Abstract: In this seminar I discuss a new and formal, measurement-driven and model-based automated verification technique, to be applied on quantitative properties over systems with partly unknown dynamics. I focus on physical systems (with spatially continuous variables, possibly noisy), driven by external inputs and accessed under noisy measurements. I formulate this new setup as a data-driven Bayesian model inference problem, formally embedded within a quantitative, model-based verification procedure.

Who: Lucas Cordeiro, University of Oxford

Title: DSSynth: An Automated Digital Controller Synthesis Tool for Physical Plants

Abstract: We present an automated MATLAB Toolbox, named DSSynth (Digital-System Synthesizer), to synthesize sound digital controllers for physical plants that are represented as linear time invariant systems with single input and output. In particular, DSSynth synthesizes digital controllers that are sound w.r.t. stability and safety specifications. DSSynth considers the complete range of approximations, including time discretization, quantization effects and finite-precision arithmetic (and its rounding errors). We demonstrate the practical value of this toolbox by automatically synthesizing stable and safe controllers for intricate physical plant models from the digital control literature. The resulting toolbox enables the application of program synthesis to real-world control engineering problems. A demonstration can be found at https://youtu.be/ hLQslRcee8.

Who: Cristina David, University of Oxford

Title: Solving Second-Order Constraints with Program Synthesis

Abstract: In this talk I’ll summarise our work on expressing and solving program analysis problems as second-order satisfiability. I’ll start by introducing a fragment of second-order logic that is expressive enough to capture numerous program analysis problems (e.g. safety proving, bug finding, termination and non-termination proving, refactoring). Subsequently, I will describe the solver we built for this fragment, which is based on program synthesis. In particular, our synthesiser is an instance of the Counterexample Guided Inductive Synthesis (CEGIS) framework and makes use of symbolic bounded model checking and genetic programming. I’ll end by discussing in more detail one of the areas where we’ve successfully applied our synthesiser, namely the synthesis of safe digital feedback controllers for physical plants represented as linear, time-invariant models.

All are welcome!

Artur Garcez

MPhil-PhD transfer presentation

Dan Philps

Thursday, 21 September 2017, 2pm

Room: A109 (College building)

Title: Motif Memory Nets

Abstract: Motifs are recurring patterns in time-series data that have the power to explain or forecast future events/data points. In the context of lifelong learning of a time-continuous dataset, “motif awareness” would improve modelling accuracy when compared to “motif naïve” generalizations. This study introduces a “motif aware” methodology called Motif Memory Nets (MMN), an ensemble approach based on a multi-column architecture of deep feedforward neural networks, employing a memory framework using Dynamic Time Warping. MMN identify, remember and recall “models” of repeating motifs in time continuous, multivariate time-series data, achieving an improvement in accuracy and, in some cases, parsimony over generalised approaches tested on the same problems. The contributions of this work are several: Firstly, this study introduces novel memory augmentation using a simple memory crystallization methodology that does not require traditional “gating”. Secondly, for memory retrieval, it is found that it is not necessary to fully learn the complex and potentially sparse distribution of the way a “memory associates with the problem space”, instead using a conditional framework drawn from the idea of “motif discovery” (or “similarity”) results in a much reduced memory retrieval and balancing task. Thirdly, MMN’s deep architecture draws on several threads in the literature applying “memory augmentation” but uses a synchronous, parallel architecture which can support distilling, while allowing MMNs to be theoretically deployable across processes and machines. Fourthly, MMNs are successfully applied to a complex, real-world dataset in the domain of Finance, a novel application for memory augmented approaches.

Keywords: Deep neural network, memory model, motif, dynamic time warping, multi-column, ensemble, time-series, lifelong learning.

Short bio: Dan Philps is a career Fund Manager and quantitative finance researcher and is working towards achieving a PhD in Machine Learning, applied (currently) to the domain of Finance, with supervisors Prof Artur Garcez and Dr Tillman Weyde. Commercially, Mr. Philps is Head of Rothko Investment Strategies – the quantitative equity investment group – and chairs the Rothko Investment and Research Committee. Prior to this, he was a Senior Fund Manager in Mondrian Investment Partners’ Global Fixed Income and Currencies team and before joining Mondrian, in 1998, Mr. Philps was a Consultant Quantitative Analyst/Programmer in the equity and derivatives businesses of Dresdner-KB, Bankers Trust and Barclays Capital, specializing in trading and risk models. Mr. Philps has a BSc (Hons) from King’s College London, is a CFA Charterholder, a member of the CFA Institute and a member of the CFA Society of the UK.

All welcome!

Machine Learning Seminar

Thursday, 1 June, 2pm, ELG10

Jakob Foerster (University of Oxford)

Title: Counterfactual Multi-Agent Policy Gradients

Abstract: Cooperative multi-agent systems can be naturally used to model many real world problems, such as network packet routing or the coordination of autonomous vehicles. There is a great need for new reinforcement learning methods that can efficiently learn decentralised policies for such systems. To this end, we propose a new multi-agent actor-critic method called counterfactual multi-agent (COMA) policy gradients. COMA uses a centralised critic to estimate the Q-function and decentralised actors to optimise the agents’ policies. In addition, to address the challenges of multi-agent credit assignment, it uses a counterfactual baseline that marginalises out a single agent’s action, while keeping the other agents’ actions fixed. COMA also uses a critic representation that allows the counterfactual base- line to be computed efficiently in a single forward pass. We evaluate COMA in the testbed of StarCraft unit micromanagement, using a decentralised variant with significant partial observability. COMA significantly improves average performance over other multi-agent actor-critic methods in this setting, and the best performing agents are competitive with state-of-the-art centralised controllers that get access to the full state.

ML seminar

Kaio Motawara

When: 15 May 2017, 1pm

Where: AG10

Abstract: Appropriate feature engineering, the process of combining and transforming raw features, can contribute significantly to improving the performance of a supervised learning task. Because of the combinatorial explosion of possible engineered features, very often successful feature engineering in a corporate environment requires domain knowledge. However, with the increasing complexity of data and numbers of sources of data, applying human insight to the feature engineering process becomes difficult.

In this presentation, I introduce a novel, fast, and hopefully elegant algorithm where the complexity of the process of feature engineering is only connected to the number of features relevant to modelling the target variable, instead of all the features in a dataset. When the number of relevant features is much smaller than the total number of features, this amounts to increasing the efficiency to almost constant time. This method draws in insights from previous work linking machine learning and information theory. I then present the results of tests of the algorithm, which show that the engineering method I have developed is indeed effective in creating a feature that improves the performance of both classification and regression algorithms when the engineered feature is included along with the rest of the features. In order to reach statistically valid conclusions, it is necessary to test the algorithm on large numbers of appropriate datasets. Therefore I also introduce a method by which to generate synthetic datasets with desirable characteristics on which to test machine learning algorithms.

At this stage, the work is a proof of concept of the method created. Future work would include creating more generalised methods and coding these into Python packages for use by the data science community.

Machine Learning seminar

Prof Duncan Gillies, Imperial College London

Monday, 20 March 2017, 2pm. College Building, room AG22

Title: Visual Cognition in Face Recognition

(work by C. E. Thomaz, V. Amaral, G. A. Giraldi, D. F. Gillies and D. Rueckert)

Abstract: Recently new methods have been emerging in the field of image recognition. In particular, deep neural networks have reached very high levels of accuracy out performing other learning paradigms on large image data bases However, although deep networks are biologically inspired they tell us very little about how human visual perception works. They can correctly classify a picture, but they cannot give us any interpretation of that picture. Human vision has been found to operate quite differently. When viewing a picture we tend to concentrate our gaze at a small number of key points rather than looking at it globally. Our efforts are directed towards interpreting rather than classifying. An interesting question is “is it possible to characterise and make use of the mechanisms of human visual perception?”

The talk will focus on face recognition, and introduce some recent work carried out as a joint project between Centro Universitario FEI and Imperial College on incorporating human visual cognition with existing face recognition algorithms. It will give a brief overview of the history of face recognition including Sirovich and Kirby’s work on face spaces using PCA and Yarbus’ eye tracking experiment on human perception. It will then go on to describe some experiments where human fixation points have been used to spatially weight the PCA face feature space leading to improved accuracy of recognition.

The 12th edition of the Workshop on Neural-Symbolic Learning and Reasoning will take place at City, University of London, on 17 and 18 July 2017.

The workshop studies the combination of well-founded symbolic AI with robust neural computing to help tackle data-driven challenges in many areas of application, from health to finance, transport and global business.

Information about paper submission deadlines, how to register, and the new industry track can be found at http://www.neural-symbolic.org/NeSy17

19 Dec 2016, AG05, 2pm-5pm

Logic Tensor Networks, Knowledge Extraction and Applications

Speakers:

Ivan Donadello (FBK, Trento)

Adriana Danilakova (City, London)

Luciano Serafini (FBK, Trento)

Edjard Mota (UFAM, Manaus and City, London)

Simon Odense (City, London)

Dan Philps (Mondrian and City, London)

All are welcome!

Artur Garcez

ps. CoCo@NIPS’16 was a great workshop: http://daselab.cs.wright.edu/nesy/CoCo2016/

Who: Dr Eric Humphrey

When: Monday 5th Dec 2016, 16:00-17:00

Where: AG08 College Building, City, University of London

Title: Open-MIC – The Open Music Instrument Classification Challenge

 Abstract: The Open-MIC initiative is a community-driven experiment to benchmarking content-based MIR algorithms in a transparent and sustainable way. Eric will present the inspiration behind Open-MIC, the approach to building an open music dataset, and the engineering effort behind making the project work. Open-MIC is focused on music, but these topics that are widely applicable to general machine learning and signal processing evaluation.

 About the speaker: Eric Humphrey is a Senior Machine Learning Researcher at Spotify, focusing on algorithms and technology related to audio experience. He obtained a BSEE at Syracuse University in 2007, an MSc in Music Engineering Technology (U of Miami) in 2009 and a PhD in Music Technology at NYU in 2015. Eric is also a multi-instrumentalist, has been a visiting lecturer at the University of Miami, worked as an independent contractor roles for several audio technology companies, spent a summer at Google doing really fun things he can’t really talk about, and currently serves as the secretary to the International Society for Music Information Retrieval.