Date: Wednesday 19th January 2011, 11:15, CSE102/3 Computer Science, Univ. of York
Speaker: Simon O'Keefe, Department of Computer Science, Univ. of York
Neural Networks: A Brief Tutorial
In this tutorial I intend to introduce the basic ideas underlying computing with neural networks. I will start with the neuron itself, and the abstract model of a neuron that underlies much of the work on artificial neural networks. I will show that even a single neuron can do useful computation, and illustrate the sorts of problems that can be solved with more complex networks of such simple neurons - Multilayer Perceptrons. I will finish with an brief discussion of some of the variants of the neuron model.
Date: Wednesday 26th January 2011, 11:15, CSE102/3 Computer Science, Univ. of York
Group meeting
Date: Wednesday 2nd February 2011, 11:15, CSE202 Computer Science, Univ. of York
Speaker: Siva Reddy, Department of Computer Science, Univ. of York
Compositionality Detection using Vector Space Model: How to distinguish "pink elephants" from "brown elephants"
Compositionality is a linguistic phenomenon where two or more words combine to form a multi-word whose meaning can be inferred from the meanings of the constituent words. But certain multi-words do not obey the compositionality principle i.e. the semantics of the multi-word cannot be interpreted from the constituent words. In this talk, I present a vector space model for determining if a multi-word is compositional or non-compositional.
Date: Wednesday 9th February 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Jim Austin, Department of Computer Science, Univ. of York
An Introduction To AURA
Date: Wednesday 16th February 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Ioannis Klapaftis, Department of Computer Science, Univ. of York
Word Sense Induction & Disambiguation Using Hierarchical Random Graphs
Graph-based methods have gained attention in many areas of Natural Language Processing (NLP) including Word Sense Disambiguation (WSD), text summarization, keyword extraction and others. Most of the work in these areas formulate their problem in a graph-based setting and apply unsupervised graph clustering to obtain a set of clusters. Recent studies suggest that graphs often exhibit a hierarchical structure that goes beyond simple flat clustering. We present an unsupervised method for inferring the hierarchical grouping of the senses of a polysemous word. The inferred hierarchical structures are applied to the problem of word sense disambiguation, where we show that our method performs significantly better than traditional graph-based methods and agglomerative clustering yielding improvements over state-of-the-art WSD systems based on sense induction.
Date: Wednesday 23rd February 2011, 11:15, CSE102 Computer Science, Univ. of York
No Meeting - Speaker cancelled due to illness
Speaker: Jelena Mirkovic, Psycholinguistics Research Group,
Department of Psychology, Univ. of York
Topic: Computational and behavioral studies of grammatical category acquisition and use
Knowledge of grammatical categories such as nouns and verbs lies at the foundations of human language comprehension and production abilities. Words' distributional and phonological properties contribute to both adult and infant learning of these categories. I will present studies investigating the contribution of semantic cues to the acquisition of grammatical categories using grammatical gender. Grammatical gender is traditionally considered a semantically arbitrary category, however there may be finer-grained correlations between semantic categories and gender classes. I will present results from a corpus analysis, an artificial neural network model and human artificial language learning experiments demonstrating the influence of semantic regularities on the acquisition and use of grammatical categories.
Date: Wednesday 2nd March 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Sam Devlin, Department of Computer Science, Univ. of York
Multi-Agent, Potential-Based Reward Shaping
Typically reinforcement learning agents are deployed with no prior knowledge despite system designers often possessing some idea of how to solve the problem. Knowledge based reinforcement learning is the study of how to incorporate these ideas into agents, allowing them to perform well in more difficult learning tasks. My interest focuses on environments with multiple reinforcement learning agents deployed, significantly changing the goals of learning and the effect of knowledge based methods. In this talk I will introduce the theory of how to correctly formalise prior knowledge for use in multi-agent reinforcement learning and demonstrate it with ample demonstrations from simple games to the complex RoboCup KeepAway Soccer Simulator.
Date: Wednesday 9th March 2011, 11:15, CSE102 Computer Science, Univ. of York
Group meeting
Date: Wednesday 16th March 2011
10:15, CSE102 Computer Science, Univ. of York
Research Students Meet and Greet: Johnathan Schaeffer
14:00, CSE082 Computer Science, Univ. of York
Speaker: Johnathan Schaeffer, University of Alberta
Topic: Six (or Less) Degrees of Separation in Social Networks
Social networks play an increasingly important role in today's society. Special characteristics of these networks make them challenging domains for the search community. In particular, social networks of users can be viewed as search graphs of nodes, where the cost of obtaining information about a node can be very high. This talk addresses the search problem of identifying the degree of separation between two users. New search techniques are introduced to provide optimal or near-optimal solutions. The experiments are performed using Twitter, and they show an improvment of several orders of magnitude over greedy approaches. Given two random users, a near-optimal solution can be found by making only an average of 9.5 requests for information to Twitter.
Date: Thursday 17th March 2011, 19:00, Ron Cooke Hub, Univ. of York
Speaker: Johnathan Schaeffer, University of Alberta
Topic: Computer (and Human) Perfection at Checkers
In 1989, the Chinook project began with the goal of winning the human World Checkers Championship. There was an imposing obstacle to success - the human champion, Marion Tinsley. Tinsley was as close to perfection at the game as was humanly possible. To be better than Tinsley, the computer had to be perfect. In effect, it had to solve checkers. Little did the project group know that their quest would take 18 years to complete. In this public lecture, Jonathan Schaeffer, the creator of Chinook, tells the story of the quest for computer and human perfection at the game of checkers.
Date: Wednesday 27th April 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Jan Staunton, Department of Computer Science, Univ. of York
Topic: Using Machine Learning to Detect Concurrent Faults
The detection of faults in concurrent software/systems has long been a topic of intense study. A number of techniques exist for detecting concurrent software, including both static and dynamic approaches. Whilst static approaches have had some success, there is still a major issue of being swamped false positives. Dynamic approaches have issues with repeatability and cost to execute. The most promising approach in recent years to detecting current faults is to use model checking. Model checking can automatically verify that a system satisfies a given specification by exhaustively checking the state space of a software program. However, model checking can fall down on large state space, and this often limits the applicability of model checking to practical industrial software.
In this presentation, I will present a machine learning based approach to detecting faults in concurrent software. Using an approach based on Estimation of Distribution Algorithms, a variant on popular evolutionary algorithms, one can focus the search of the state space on areas that are most likely to contain an error. I will describe the technique, and provide empirical evidence demonstrating its potential.
Date: Wednesday 11th May 2011, 11:15, CSE102 Computer Science, Univ. of York
Speakers: Chris Bak and Andrew Owenson
Undergraduate Student Project Talks
Date: Wednesday 18thth May 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Jelena Mirkovic, Psycholinguistics Research Group, Department of Psychology, Univ. of York
Topic: Computational and behavioral studies of grammatical category acquisition and use
Knowledge of grammatical categories such as nouns and verbs lies at the foundations of human language comprehension and production abilities. Words' distributional and phonological properties contribute to both adult and infant learning of these categories. I will present studies investigating the contribution of semantic cues to the acquisition of grammatical categories using grammatical gender. Grammatical gender is traditionally considered a semantically arbitrary category, however there may be finer-grained correlations between semantic categories and gender classes. I will present results from a corpus analysis, an artificial neural network model and human artificial language learning experiments demonstrating the influence of semantic regularities on the acquisition and use of grammatical categories.
Date: Wednesday 1st June 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: James Cussens, Department of Computer Science, Univ. of York
Topic: Bayesian Network Learning with Cutting Planes
The problem of learning the structure of Bayesian networks from complete discrete data with a limit on parent set size is considered. Learning is cast explicitly as an optimisation problem where the goal is to find a BN structure which maximises marginal likelihood (BDe score). Integer programming, specifically the SCIP framework, is used to solve this optimisation problem. Acyclicity constraints are added to the integer program (IP) during solving in the form of *cutting planes*. Finding good cutting planes is the key to the success of the approach---the search for such cutting planes is effected using a sub-IP. Results are good with optimal BNs being found substantially faster than competing approaches. I will also discuss whether 'variable pricing' can be used to remove the parent set size limitation.
Date: Wednesday 22nd June 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Richard Gil Herrera, University of Granada, Spain
Topic: Knowledge Engineering Supported By A Systemic Ontology Learning Approach
Companies and organizations are demanding an efficacious knowledge management to face the challenges of today's modern world. For that reason, they are increasing system innovation investments to turn information into useful knowledge for decision making obtained from heterogeneous Knowledge Sources (KSOs) such as databases, documents, and even ontologies developed previously.
Methodological Resources (MRs) such as tools, methods, and techniques for the required knowledge discovering and recovering purposes have gradually become more elaborated and mature in the framework of Knowledge Engineering. Particularly, in the Ontology Learning (OL) field, these MRs are usually related to one specific KSO, in an isolated way and without any relation among the others KSOs. OL methodologies under a systemic perspective could be taking some advantage of those MRs, and they may favourably impact the total quality performance of the associated Knowledge Management Systems.
The main contributions provided by this work are on the one hand, a novel user-centred Systemic Methodology for Ontology Learning (SMOL) from heterogeneous KSOs, and a proposal of an OL Knowledge Support System (OLeKSS) model. It is applied for an academic study case until a whole knowledge acquisition process is completed. On the other hand, a methodological evaluation of SMOL under technical methods based on feature analysis is developed.
An open problematic about the semantic quality of the ontology updated and enriched during those OL processes is the main concern with this stay research. For this case study is trying to take advantage of the novel methodologies of word sense induction and disambiguation (developed in UoY) to improve the host-ontology’s semantic quality from a domain-customized corpus.
Date: Wednesday 29th June 2011, 11:15, CSE102 Computer Science, Univ. of York
1st Speaker: Alan Frisch, Department of Computer Science, Univ. of York
Topic: Relaxing the Assumptions for Applying Local Search
Though local search is one of the most general techniques used in artificial intelligence, certain assumptions are made about the domains in which it is implied. One assumption is that states can be evaluated quickly, which enables local search to examine large numbers of states. Another assumption is that the objective value of a state can be determined accurately, which means that there is no need to perform multiple evaluations of a given state. This talk asks how local search should be adapted to work in domains where these two assumptions are not met. Please rest assured that no answers will be provided.
2nd Speaker: Daniel Kudenko, Department of Computer Science, Univ. of York
Topic: Knowledge-Based Reinforcement Learning (KBRL)
Reinforcement learning (RL) is a highly popular machine learning technique, mainly due to its natural fit to the agent paradigm (i.e. learning by repeatedly acting and sensing in an environment) and its resulting wide application potential.Despite these advantages, RL suffers from scalability problems which have prevented its successful use in many complex real-world domains. The KBRL approach is focused on the use of domain knowledge to scale-up and improve reinforcement learning and support transfer learning. Conversely, reinforcement learning can also be used to revise domain knowledge. In my talk I will give a brief overview of KBRL, the research done so far, and current and future directions.
Date: Monday 11th July 2011, 13:00, CSE102 Computer Science, Univ. of York
Speaker: Waleed Alsanie, Department of Computer Science, Univ. of York
Topic: Learning PRISM Programs for Generative Modelling (Thesis Seminar)
Clausal logic is a very rich relational representation which has been used to represent knowledge in AI applications. However, the fact that in clausal logic, clauses are evaluates to crisp true/false values has limited its use to represent different problem where uncertainty needs to be encoded. In this talk, I will introduce a formalism called PRogramming In Statistical Modelling (PRISM) which combines both probabilistic modelling and logic programming in one framework. I will then present thework that has been done so far to build on existing two sub-areas of machine learning Inductive Logic Programming (ILP) and learning probabilistic models to learn generative PRISM programs from examples and Background Knowledge (BK).
Date: Wednesday 26th October 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Jing Kan, Department of Computer Science, Univ. of York
Topic: Spatial-temporal Source Reconstruction for MEG
Magnetoencephalography (MEG) is a new non-invasive technique for functional imaging of the human brain, which has been widely used in both research and clinical application. The technique is to measure the magnetic field surrounding head that via the extremely sensitive sensors located outside of scalp. The measured magnetic field is mainly generated by the electronic activity in brain. Compared with fMRI, MEG presents superior temporal resolution, which complements the weakness of brain imaging in the time domain. Technically, MEG source reconstruction is an ill-posed inverse problem. The main current solutions have their intrinsic weakness, which provides the potential space for exploring a new possible solution by applying the knowledge of the new research field, such as pattern recognition.
Our work makes the connection between the field of classical pattern recognition, graph theory and MEG source reconstruction. The whole design of the algorithms is based on the MEG spatial reconstruction at a single time point. It is assumed that the 3D sources distributed on each vertex of the original cortical surface generated by structural MRI of the same subject. Then, the basis function algorithm is designed to spatially reconstruct the source distribution at the specific time point. Subsequently, another method from pattern-recognition, Bayesian super-resolution algorithm, is introduced to expand the reconstructed source distribution from the original mesh into the interpolated high-resolution mesh, through the process of which the spatial resolution of the reconstruction is developed. Furthermore, as the MEG measurement system is assumed to be a linear dynamic system, one of the classical solutions, Kalman smoother, is finally used to improve the temporal resolution of this source distribution based on the high-resolution mesh.
In summary, our work combines multiple classical methods of pattern recognition with the MEG spatial-temporal source reconstruction in order to achieve a highly sensitive spatial and temporal reconstruction. This combination opens a new window for the MEG source reconstruction problem with a novel angle, and provides new possibility for the further MEG research on reconstruction solution.
Date: Wednesday 2nd November 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Ali Abusnina, Department of Computer Science, Univ. of York
Topic: Soft Sensors
In striving towards computing automation, different industries in different applications are seeking effective tools to monitor and control increasingly complicated processes; from food processing and paper industry to nuclear plants and chemical refineries. Such tools are soft sensors which are gaining wide popularity as the technology advances for their increasing advantages.
Soft sensors are inferential estimators that draw conclusions from process observations when hardware sensors are unavailable or unsuitable for cost or reliability reasons; they have an important auxiliary role in sensor validation when performance declines through senescence or fault accumulation.
Date: Wednesday 9th November 2011, 14:00, CSE082 Computer Science, Univ. of York
Distinguished International Speaker: Michele Sebag, Universite de Paris-Sud
Host: Dimitar Kazakov
Topic: New advances in Artificial Intelligence and Machine Learning: Playing Go at expert level; lessons learned
Optimization in front of uncertainty and sequential decision making come in many flavors in computer science. Games provide good testbeds to study both issues, as they explore closed worlds with simple rules while being challenging for humans and machines. Among games, the game of Go succeeded the game of chess as prominent challenge for artificial intelligence since 1997 (Deep Blue vs Kasparov match). The talk will describe the advances in computer-Go achieved in the last five years. These rely on new tree search algorithms, rooted in Multi-Arm Bandit algorithms. If time permits, we shall discuss how this very approach can be adapted to other sequential decision problems, e.g. optimal energy policy.
Date: Wednesday 16th November 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Matthew Butler, Department of Computer Science, Univ. of York
Topic: Implications of the Adaptive Market Hypothesis for Forecasting
This talk covers research concerning the validity and implications of the Adaptive Market Hypothesis (AMH) from a computational intelligence (CI) perspective. The AMH is a relatively new theory of financial market behaviour that has several implications which are in contrast, to the widely held, Efficient Market Hypothesis. Among these implications are variable market efficiency, cyclical profitability and path-dependence. Each of these will be addressed from a computational intelligence perspective which endeavours to evaluate the validity and then create forecasting tools inspired from them.
Date: Wednesday 23th November 2011, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Sam Devlin & Kyriakos Efthymiadis, Department of Computer Science, Univ. of York
Topic: Plan-Based Reward Shaping and Belief Revision for Multi-Agent Reinforcement Learning
Our recent theoretical results have justified the use of potential-based reward shaping as a way to improve the performance of multi-agent reinforcement learning (MARL). However, the question remains of how to generate a useful potential function. Previous research demonstrated the use of STRIPS operator knowledge to automatically generate a potential function for single-agent reinforcement learning. Following up on this work, we investigate the use of STRIPS planning knowledge in the context of MARL. Our results show that a potential function based on joint or individual plan knowledge can significantly improve MARL performance compared with no shaping. In addition, we investigate the limitations of individual plan knowledge as a source of reward shaping in cases where the combination of individual agent plans causes conflict.
A possible solution to these limitations is using Belief Revision in order to update an agent's knowledge base. Faulty knowledge can slow down an agent's learning process significantly, or inhibit learning alltogether. Belief revision focuses on updating a knowledge base when new information arises, so that it remains consistent. With the agent acting constantly within its environment, faulty knowledge can be identified and corrected in order to produce more accurate plans and thus improve an agent's learning speed and policy quality.
Date: Wednesday 30th November 2011, 11:15, CSE102 Computer Science, Univ. of York
No Meeting - Strike day
Speaker: Tommy Yuan, Department of Computer Science, Univ. of York
Topic: Argument-based approach to computer system safety engineering
Safety case development is not a post-development activity, rather it should occur throughout the system development lifecycle. The key components in a safety case are safety arguments. Too often, safety arguments are constructed without proper reasoning. Inappropriate reasoning in safety arguments could undermine a system's safety claims, which in turn contributes to safety-related failures of the system. To address this, we argue that informal logic argument schemes have important roles to play in safety arguments construction and review process. Ten commonly used reasoning schemes in computer system safety domain are proposed against the safety engineering literature. The role of informal logic dialogue games in computer system safety arguments reviewing is also discussed and a dialectical model for safety argument review is proposed. It is anticipated that this work will contribute toward the development of computer system safety arguments, and help to move forward the interplay between research in informal logic and research in computer system safety engineering.
Date: Wednesday 7th December 2011, 11:15, CSE102 Computer Science, Univ. of York
Group Meeting
Date: Wednesday 14th December 2011, 11:15, CSE102 Computer Science, Univ. of York
First Year RSs Literature Review
Peter Scopes (11:15-11:45):
MARL - a brief history
I'll be focusing on the history and development of Multi Agent Reinforcement Learning.
Dorothy Thato Kentse (11:45-12:15):
Reinforcement learning for data mining