Distinguished Lectures

In the first part we give an overview of the whole-genome shotgun sequencing method with paired end-reads that we developed and used to sequence Drosophila (2000), Human (2001) and Mouse (2001), in quick succession. Because the finishing phase of genome assembly is an order of magnitude more expensive then the shotgun phase, most genomes being sequenced today will never be finished. This makes the goal of building better whole genome assemblers more important than ever. A string graph concept developed by us (2005) portends near perfect assembly and has provided the core idea for new advances. We present the history and ideas in brief.

The second part of the talk focuses on our new work on the development of algorithms and software for the automatic interpretation of images produced by light and electron microscopy of stained samples with a particular emphasis on building 3D and 4D "atlases" of brains, developing organisms, and cellular processes. Specific project include (a) the in vivo monitoring of mitotic processes in a developing C. elegans embryo (with T. Hyman, MPI-CBG, Dresden), (b) the in situ observation of C. elegans' cells in stages L1 through adult to decode their transcriptional and functional states (with S. Kim, Stanford), (c) the detailed neuro-anatomy of an entire fly brain including all synaptic connections via a hybrid approach combining EM and light-microscopy (with colleagues at Janelia), and (d) mapping the developmental trajectory of the fly brain via the visual analysis of MARCM clones and promotor fusions (with V. Hartenstein, UCLA, G. Rubin, J. Simpson, and J. Truman, JFRC).

This talk explores some of the differences between pure and applied research by comparing two AI projects undertaken at Thomson Corporation.  One, an editorial tool called History Assistant based on information extraction technology, was an academic success resulting in multiple publications but deemed a failure internally.  The other, a recommender system called ResultsPlus that used multiple kinds of machine learning, was more or less unpublishable but created a $40M/year business.  In addition to technology issues, the talk addresses the role of luck, people, and timing in the successful application of advanced technologies.

Lecture slides: ppt, pdf

Peter Jackson holds a Ph.D. in Artificial Intelligence from the University of Leeds and taught in the Department of Artificial Intelligence at Edinburgh University from 1983 to 1988.  In 1995, he joined the Thomson Corporation, where he is now Chief Scientist and vice president of R&D.  He runs a group of over 40 researchers who work with business units in the legal, financial, science and healthcare sectors to deliver custom information solutions.  His latest book, “Natural Language Processing for Online Applications”, came out in a 2nd edition in 2007.

One major difference between human tutors and current computer tutors is that only human tutors participate in unconstrained natural language dialogue with students.  This difference has led to the conjecture that human tutoring might be so effective because of its use of natural language dialogue.  Potential advantages of natural language dialogue as a learning environment include providing opportunities for both the tutor to infer information about a student, and the student to participate more actively in the learning process. Thus in recent years, the development of automated tutorial dialogue systems has emerged as an important topic of research in the field of technology enhanced learning.

What difference does it make to a market for there to be a well-regarded mathematical model of the market, especially one that is not just an external analysis by academics but is used by market practitioners?

This talk will ask this question mainly in regard to the most famous model in modern financial economics, the Nobel-Prize winning Black-Scholes-Merton model of option pricing, which is the core mathematical foundation of the global market in ‘financial derivatives’. (At the end of June 2006, derivatives contracts outstanding worldwide totaled $454 trillion, the equivalent of nearly $70,000 for every human being on Earth.)

The talk will describe how the practical uses of the model initially had the effect of making markets more like the postulates of the model, but will discuss how this effect reversed in direction in the 1987 stock market crash, with near-disastrous consequences for the global financial system.

No previous knowledge of economics will be necessary to understand the talk: what an ‘option’ and a ‘derivative’ are, and what the Black-Scholes-Merton model consists in, will be explained in simple terms.

Using the achievements of my research group over the last 35+ years, I provide evidence to support the following hypothesis: By complementing each other, cooperating reasoning process can achieve much more than they could if they only acted individually.

Alan Bundy is the only researcher to have won both the IJCAI Award for Research Excellence and the IJCAI Distinguished Service Award.  This talk is based on his acceptance speech for the Research Excellence Award, delivered at IJCAI 2007.

Computer Science research in the area of the World Wide Web has largely focused on improved search for individual web pages or on the modeling of Web connectivity (using the tools of networking). However, given the huge impact of the Web on our world, this seems to be an impoverished view. What are the principles of engineering that have made the Web flourish? How can we engineer new technologies, that will extend the capabilities of the Web? What are the social impacts of Web use, and how can Web technologies both allow greater freedoms while preserving the ones we have?

In this talk, I will use some examples from Semantic Web and "policy aware" information access to demonstrate new Web technologies and how we might explore some of the trade-offs between making it easier to integrate information on the Web with protecting that information from abuse. I will explore some of the emerging trends on the Web including social networking, blogging, and beyond page search, and discuss some of the research and technology challenges that they pose to continued Web growth and access, and some new technologies being explored to address these challenges.

Download the slides

In the last 10 years, the combination of techniques from machine learning and operations research has allowed major advances in learning and planning for uncertain environments. Reasonably large problems can be solved using current techniques. But what if we want to scale up to the uncertain learning and planning problem that you face every day? It is many orders of magnitude larger than the biggest problem we can solve currently.

Over the last six years, there has been a surge of interest in the economics of system security. It's become clear that many systems fail not because of poor design decisions, but because incentives were wrong. For example, the organisation that protects a system may not suffer the full costs of failure - and may even be using the system to transfer liability to others. Microeconomic analysis can be just as important as cryptanalysis. The interplay between policy and technology is rapidly spreading to other fields: the debate on the relative merit of open and closed systems, the management of the patching cycle, digital rights management, and the causes of large-scale systems failure are just four of the topics I find interesting. I will discuss recent research, and then ask what we might expect in a future world of pervasive computing - a world in which most devices come to contain CPUs and communications, in which more and more value is added by software, and in which more and more industries come to resemble the software industry.

The talk will discuss the idea of comprehensive and realistic modeling of biological systems. In comprehensive modeling the main purpose is to understand an entire system in detail and to use that understanding to analyze and predict behavior in silico. In realistic modeling the main issue is to model the behavior of actual elements, making possible realistic and executions/simulations that reveal emergent properties. I will address the motivation for such modeling and the philosophy underlying the techniques for carrying it out, as well as the crucial question of when such models are to be deemed valid, or complete. The examples I will present will be from among the modeling efforts I have been involved in with my group: T cell development in the thymus, lymph node development, the pancreatic islets, and the C. elegans reproduction system.

Current research on distributed knowledge processes suggests a critical conflict between knowledge processes in groups and the technologies built to support them. The conflict centres on observations that authentic and efficient knowledge creation and sharing is deeply embedded in an interpersonal face to face context, but that technologies to support distributed knowledge processes rely on the assumption that knowledge can be made mobile outside these specific contexts.  This paper draws on the history of the development of large scale information infrastructures over the past millennium to explore the mutual imbrication of knowledge, organizational form and political agenda and adduce some lessons for the contemporary push to create e-science.

Richard Stallman talked about the goals and philosophy of the Free Software Movement, and the status and history the GNU operating system, which in combination with the kernel Linux is now used by tens of millions of users world-wide.

About Richard Stallman

Richard Stallman is the founder of the GNU Project, launched in 1984 to develop the free software operating system GNU. The name 'GNU' is a recursive acronym for 'GNU's Not Unix'.

Stallman graduated from Harvard in 1974 with a BA in physics. During his college years, he also worked as a staff hacker at the MIT Artificial Intelligence Lab, learning operating system development by doing it. He wrote the first extensible Emacs text editor there in 1975. He also developed the AI technique of dependency-directed backtracking, also known as truth maintenance. In January 1984 he resigned from MIT to start the GNU project.

Since the origins of what we now call computer science, the observed or speculated behaviour of the nervous system has been a source of inspiration for computational models. We will present several inter-related mathematical models whose simplest version is the "spiked random neural network model" which we described a decade ago, while more complex versions allow us to model genetic computation or the behaviour of adversarial agent populations. We will briefly discuss the mathematical properties of these models, including stationary distributions, the existence and uniqueness of solutions, and their use as universal approximators of continuous functions. We will then discuss how these models can be used to study, in relation to electrophysiological measurements from simultaneous multiple cell recordings in natural neuronal sub-assemblies, certain significant properties of the somatosensory system. Finally, we will illustrate the models' role in engineering applications by showing how they can be used to process texture in images, or to control packet flow in a network. The lecture will be based on our published research and patents, in collaboration with several post-graduate students.

Many programs traverse data structures built from rich mutually-recursive data types. Such programs often have a great deal of "boilerplate" code that simply walks the structure, hiding a small amount of "real" code that constitutes the reason for the traversal. "Generic programming" is the umbrella term to describe a wide variety of programming technology directed at this problem. All these techniques aim to provide mechanical support for the "boilerplate" part, leaving the programmer free to concentrate on the important part of the algorithm. Such generic programs are much more robust to data structure evolution because they contain many fewer lines of type-specific code.

The trouble is that most generic programming techniques either require significant support from the programming language itself, or are inconvenient for the programmer. In this talk I will describe a new approach to generic programming based on the functional language Haskell, which combines programming convenience with very modest demands on the language. Our approach is simple to understand, elegant, reasonably efficient, and handles all the data types found in conventional functional programming languages. It makes essential use of rank-2 polymorphism, an extension found in some implementations of Haskell.

The talk is a development of work first reported in "Scrap your boilerplate", Laemmel & Peyton Jones, Proc ACM SIGPLAN Workshop on Types in Language Design and Implementation (TLDI 2003), New Orleans, Jan 2003.

The human genome has been called the "book of life," a natural extension of the long-standing metaphor of DNA as a language. Taking this conceit seriously, we can ask to what extent the genome may profitably be viewed as a work of literature, subject to critical exegesis. While seemingly at opposite poles from the "hard science" of molecular biology, in fact such an approach is not so far from the increasingly hermeneutic role of the bioinformatician, insofar as both are concerned with comparing texts, detecting subtle patterns and relationships, elucidating theme and variation, etc. In this talk I will explore literary and linguistic aspects of the genome, by means of a "genomic" textual analysis of Lewis Carroll's Jabberwocky.

Freenet is free software which lets you publish and obtain information on the Internet without fear of censorship. To achieve this freedom, the network is entirely decentralized and publishers and consumers of information are anonymous. Without anonymity there can never be true freedom of speech, and without decentralization the network will be vulnerable to attack.

Communications by Freenet nodes are encrypted and are "routed-through" other nodes to make it extremely difficult to determine who is requesting the information and what its content is. Users contribute to the network by giving bandwidth and a portion of their hard drive (called the "data store") for storing files.

Unlike other peer-to-peer file sharing networks, Freenet does not let the user control what is stored in the data store. Instead, files are kept or deleted depending on how popular they are, with the least popular being discarded to make way for newer or more popular content. Files in the data store are encrypted to reduce the likelihood of prosecution by persons wishing to censor Freenet content.

This talk will explore the Freenet architecture, its underlying theory, the lessons learned from its actual implementation, and what the future holds.

Approximability of the Edge Disjoint Paths Problem

Given a graph and a set of pairs of vertices, the edge disjoint paths problem (EDP) is to maximize the number of pairs that a can be connected by edge disjoint paths. EDP and its variations capture a number of fundamental problems in combinatorial optimization and have a variety of applications in network design and routing, VLSI routing, and resource allocation. The study of EDP and related problems has been of great importance in algorithmic development. Even simple cases of EDP are NP-hard and thus much work has focused on approximation algorithms that run in polynomial time and obtain solutions that have a worst case guarantee on the quality of their output. I will talk about some of my recent work on understanding the approximability of EDP and its variations.

Phase transitions are familiar phenomena in physical systems.  But they also occur in random versions of combinatorial models, including random versions of some of the canonical problems of theoretical computer science.  In this talk, I will illustrate this by discussing joint work with Christian Borgs, Stephan Mertens and Boris Pittel on the so-called random optimum partitioning or load balancing problem -- a fundamental problem in combinatorial optimization.  I will show how this problem undergoes a phase transition from a phase in which it is typically possible to balance loads to a phase in which such balance cannot be achieved. I will also discuss how notions of phase transitions may help us to understand what makes hard problems hard.  No previous knowledge of phase transitions will be assumed in this talk.

In early 1973, Ken Thompson wrote "grep" at the request of Doug McIlroy, who needed a convenient tool to implement phonetic rules for a speech synthesizer. Almost thirty years later, finite-state methods still dominate practical text and speech processing, and also play a central role in biological sequence analysis. I will discuss several advances in finite-state techniques that have contributed to this remarkable longevity, with particular focus on probabilistic extensions used in speech recognition and information extraction. I will conclude with a brief discussion of whether we are ready to move beyond finite state.

Formal methods originally focussed on "high-end" purposes and technology, such as proofs of correctness supported by interactive theorem proving.  Later, debugging and model checking demonstrated the value and practicality of lesser ambitions and greater automation. Recently, even more modest goals, such as test-case generation and extended typechecking, are attracting attention and hold the promise that formal methods may soon penetrate mainstream engineering as "invisible" adjuncts to compilers, CAD suites, and the ubiquitous Matlab.  Paradoxically, the complete automation that is required for such "low-end" purposes demands significant advances in supporting technology. 

I will describe how mechanized abstraction can be used to support a spectrum of purposes from debugging to correctness in a context that allows greater automation than previous methods of formal analysis. I will then suggest how the spectrum can be extended to truly invisible formal methods and will outline some new techniques that provide the requisite automation.

The COOL Project (Comprehensive Object-Oriented Learning) is a 3-year research project proposal launched by by a consortium of four Norwegian research institutions, supported by research institutions in Aarhus in Denmark, and o-operating with test sites around the world. COOL will contribute to a unifying process- and object-oriented platform for informatics, and produce a "Learning Landscape" of pedagogical and organisational components to be used in a modern and system-oriented education in informatics and related  fields. It will provide an alternative to the current pedagogical approach used, commonly regarded as unsuccessful. COOL will co-operate with research institutions in Denmark and with a number of test sites (universities and colleges) around the world, representing a number of language/cultural worlds (Spanish/South American, English/North American, Scandinavian, and perhaps others). COOL will produce an introductory course, supported by a textbook and DVD records containing integrated multimedia material. The COOL Learning Landscape shall allow for alternative courses, adapted to local cultures and conditions.

Kristen Nygaard, together with Ole-Johan Dahl, is the designer  of SIMULA I (1961-65) and SIMULA 67 - the first object oriented  programming languages, which introduced the concepts upon which all  later object-oriented programming languages are built. For this work he was given the Turing Award and the John von Neumann Medal in 2001. He has also done research for Norwegian trade unions on  planning, control, and data processing, all evaluated in light of the  objectives of organised labour. He has been professor in Aarhus, Denmark  and Oslo. His work there has included research and education in system  development and the social impact of computer technology, and has become the foundation of what today is called "the Scandinavian School in  System Development", closely linked to the field of Participatory Design. Kristen Nygaard's current interests are studies of the didactical  aspects of introductory teaching of programming, and the creation  of a process-oriented conceptual platform for informatics.

We present a knowledge-based method for the definition of HTML/XML wrappers and a fully visual interface for the generation of wrappers based on the that method. We also demonstrate the LIXTO tool, which implements the proposed techniques. LIXTO is portable (implemented with Java), offers a capacious interactive visual interface, allows for expressive and flexible data extraction and uses intuitive hierarchical extraction, as well as string extraction techniques. LIXTO translates relevant parts of web-pages into XML. It can be used to create an "XML-Companion" for a HTML web page with changing content, containing the continually updated XML-translation of the relevant information. Joint work with Robert Baumgartner and Sergio Flesca. Papers on LIXTO can be found at lixto.com or at http://www.dbai.tuwien.ac.at/proj/lixto/download.html
 

In both eBusiness and eScience, we often need to integrate services across distributed, heterogeneous, dynamic ``virtual organizations'' (VOs) formed from the disparate resources within a single enterprise and/or via external resources sharing and service provider relationships. This integration can be technically challenging due to the need to achieve various qualities of service (QoS) when running on top of different native platforms. We present an Open Grid Services Architecture that addresses these challenges. Building on concepts and technologies from the Grid and Web services communities, the architecture defines a uniform exposed service semantics (the Grid service); defines standard mechanisms for creating and discovering transient Grid service instances; provides location transparency and multiple binding protocols for service instances; and supports mapping services for integration with underlying native platforms facilities. The Open Grid Service Architecture defines, in terms of Web Services Description Languages (WSDL) interfaces, mechanisms required for creating and composing sophisticated distributed systems, including lifetime management, reliable remote invocation, change management, credential management, and notification. Our presentation  describes how Globus Toolkit mechanisms can be used to implement a service oriented architecture, explaining how Grid functionality can be incorporated into a Web services framework, and illustrating how our architecture can be applied  within commercial computing as a basis for distributed system integration within and across organizational domains.

"Grid" computing has emerged as an important new field, distinguished from conventional distributed computing by its focus on large-scale resource sharing, innovative applications, and, in some cases, high-performance orientation.  In this talk, I define this new field. First, I review the "Grid problem," which I define as flexible, secure, coordinated resource sharing among dynamic collections of individuals, institutions, and resources--what I refer to as virtual organizations. In such settings, we encounter unique authentication, authorization, resource access, resource discovery, and other challenges.  It is this class of problem that is addressed by Grid technologies.  I present an extensible and open Grid architecture, in which protocols, services, application programming interfaces, and software development kits are categorized according to their roles in enabling resource sharing. I also review major Grid projects worldwide and describe how they are contributing to the realization of this architecture.

Dr. Ian Foster is Senior Scientist and Associate Director of the Mathematics and Computer Science Division at Argonne National Laboratory, Professor of Computer Science at the University of Chicago, and Senior Fellow in the Argonne/U.Chicago Computation Institute. He currently co-leads the Globus project with Dr. Carl Kesselman of USC/ISI as well as a number of other major Grid initiatives, including the DOE-funded Earth System Grid and the NSF-funded GriPhyN and GRIDS Center projects.  He founded the Global Grid Forum in 1998, and recently co-edited the book ``The Grid: Blueprint for a New Computing Infrastructure."

A number of management information systems that are aimed at the print industry include 'scheduling modules' which are intended to automate the work of scheduling print jobs to machines. These model the work of scheduling as 'rote work', involving the simple allocation of resources to jobs. Although MIS systems in general have been widely embraced in the print industry there has, however, been little take-up of the automated scheduling modules. Based upon fieldwork in over twenty print factories it is suggested that one of the reasons for this is  that the view that computational activities undertaken by people is rote and can better be done by machines misses what, in organisational or work terms, these computations may be doing. Through an examination of the work of production scheduling it will be suggested that scheduling computations are not only about the allocation of resources, but also about the construction of stable organisational structures for product delivery. In this respect, production scheduling does not so much control production but rather makes manifest the interrelationships between the decisions of the different parties to print work acquisition and delivery. Production schedulers thus work a production calculus as an aid to organisational decision making. An understanding of the organisational aspects of working the production calculus suggests that a light weight approach, in which tools to aid and support the decision making are developed, may be preferable to automation.

This talk will discuss the subset of "neural" hardware techniques that use spiking, or pulsed behaviour. The methods behind pulsed VLSI& will be reviewed and some exemplar circuits presented - for spike-generation, synaptic "multiplication" and learning. The potential and limitations  of pulsed VLSI will be revisited in the context of the new medium that is Deep-Sub-Micron (DSM) silicon.  By 2010, transistors will be smaller than 50nm - noise will be greatly amplified and irreducible. At present, no strategy exists for reliable computation with such devices. Some preliminary work in (a) probabilistic hardware and (b) spike-based learning will be discussed.

The problem of analysing visual motion, especially against dense background clutter, is challenging. Uncertainty in the positions of visually sensed features and ambiguity arising from the clutter, call for a probabilistic treatment. A promising approach is to attempt to describe the image formation process probabilistically, and then look for suitable inference engines to do the analysis. (Statistical fitering and expectation maximisation have both been tried, with some success.)

An outstanding problem is to define image features whose mutual probabilistic dependence is known -- it is impossible to set up a convincing framework for inference without doing this. In this talk we look at some possibilities, including the pixels themselves and the outputs of banks of filters such as wavelets. A new approach -- the "metric mixture" model -- is proposed to circumvent the problem. Its power for practical analysis will be illustrated on video clips of a human body moving and on changing facial expressions.

Computer systems in which autonomous software agents negotiate with one another in order to come to mutually acceptable agreements are likely to become pervasive in the next gereration, In such systems, the agents will be required to participate in a range of negotiation scenarios and exhibit a range of negotiation behaviours (depending on the context).

To this end, this talk explores the issues involved in designing and implementating a number of automated negotiators for real-world electronic commerce applications.

In this presentation, motor skills in playing the cello are extensively investigated. Virtuosos show incredible skills in performing musical instruments. This research aims to reveal the origin of such skills. 

Through observations of typical faults which amateur players apt to commit, we derived a conjecture that the most of the faults come from bad usage of muscles which causes mutual intereference of muscles. The conjecture can explain almost all faults amateur players easily commit and also it suggests how to avoid them. From the conjecture, we conclude that virtuosos have skills for avoiding such faults and we assume such skills come from satisfying a set of constraints derived by mutual interference of muscles. We propose a possible method to automatically extract these constraints from virtuosos' biomechanical peformance data by using modern machine learning methods such as inductive logic programming. We show some experimental results so far and state future work necessary for verbalization of tacit knowledge for motor skill.

Object vision depends on knowledge of objects. How much can vision itself add to the knowledge? We will discuss kinds of knowledge, how they are gained, and something of relations between perceptions and conceptions.

About Richard Gregory

In automated verification one uses algorithmic techniques to establish the correctness of the design with respect to a given property. Automated verification is based on a small number of key algorithmic ideas, tying together graph theory, automata theory, and logic.

In this self-contained talk I will describe how this "holy trinity" gave rise to automated-verification tools.

The production and comprehension of meaningful human language by machine is an as yet unattained goal whose importance is increasing as computers become cheaper, more powerful, more widespread, and more connected. Many applications that will immeasurably enhance everyday life in the information age are waiting in the wings.

Until now, the rapid progress that has generated the current inadequate technology for interacting with computers using human language has been mainly fueled by Moore's law. The exponential increase in power of computing machinery over time has meant that some very simple mechanisms which we know to be quite limited in their capabilities relative to human language processors have made most of the running, and underlie everyday applications of natural language technology in speech processing and web-based document retrieval.

Recently, however, there have been signs of a reintegration of linguistically informed models of language with probabilistic methods. Moreover, physics itself tells us that Moore's law will eventually cease to apply, and that it may do so quite soon. On both counts it is likely that further advance in natural language technology will depend on devising quite new kinds of mechanism, more closely related to human language understanding. Researching this question, and training the new generation of scientists to undertake this work, is one of the greatest scientific challenges that faces universities today, involving linguists and psychologists as well as computer scientists.

Progress on this question requires us to recognise that human utterance and understanding is, like all aspects of our being an essentially dynamic process, unfolding in time like a computation expressed in a programming language. The lecture argues that a simpler and more explanatory account of natural grammars can be achieved by abandoning standard linguistic formalisms and instead applying the formal tools that have been used to define structure and meaning for computer programming languages. Viewing grammar in this way allows us to capture many phenomena that have resisted analysis in traditional terms and are relevant to practical applications. An example is spoken intonation, the "tunes" by which speakers of English mark emphasis and relation to conversational context. The theory exhibits some very striking parallels between the linguistic apparatus and more generally applicable cognitive mechanisms and what we know about their location in the human brain.

I became involved in the design of a successor of Algol 60 in the years 1962-67. The result was Algol-W (66), and later the Algol-style Pascal (70), Modula-2 (79), and Oberon (88). In turn, they introduced the concepts of data structuring and typing, modular decomposition and separate compilation, and object-orientation.

In this talk, we summarize these developments and recount some of the influences and events that determined the design and implementation of these languages.

Throughout the ages it has been helpful to explain the workings of the human brain and, indeed, the mystery of the human brain, in terms of the current most advanced technology.

Although the human brain has been likened to a computer since the 1950s, and although there are some similarities in the two systems, the differences are far more basic.

We shall explore in this talk how the biological brain functions in a non-algorithmic way and how it is underpinned by chemical processes which could not be produced in silicon, and how we are feeling organisms as opposed to living machines.

We shall go on to explore the future interfaces between carbon and silicon systems and contemplate the ultimate prospect of a conscious computer.

Traditionally computers and computer software are regarded as deterministic, with logic providing the mathematical basis for their development. Increasingly, however, we are interested in highly complex tasks in which uncertainty prevails. Examples include novel computer input methods such as vision, speech and handwriting; intelligent interfaces which attempt to model and forecast user actions; and systems for the retrieval and interpretation of information from large data sets such as the world-wide web.

The complexity of such problems precludes conventional approaches based on hand-crafted algorithms. Instead, adaptive computation aims to learn a solution to the problem from a set of training data. Probability theory offers a consistent mathematical framework for quantifying uncertainty, and provides the theoretical basis for the development of novel adaptive computer systems. Learning itself can be viewed as a reduction in uncertainty, and is quantified through Bayes' theorem.

In this talk I will describe a number of recent developments in adaptive computation, and highlight some of the current research issues. I will also explore the potential impact which this research may have on future computer technology.

The electronic computer bestrode the world stage in the second half of the 20th century, moving from a status in 1943 when the president of IBM predicted that "there is a world market for about five computers" to a status in 1999 when, for example, 30% of UK households had home computers. Over this period, the cost, size, weight and power consumption of computers shrank dramatically, while their speed and storage capacity expanded equally dramatically. However, architecturally, the typical computer of 1999 is not radically different from a computer of 1949. The situation is set to change in the near future. This lecture explores what is in prospect, both for the computer user and for the computer architect. One question raised is whether the term "computer" itself will continue to be apt.

For the computer user, the terms "hard" and "soft" here have meanings related to usability. Undoubtedly, computers became more user-friendly over the past 50 years, as the user community broadened from technical specialists to the general public. However, they are still relatively hard to master. One consequence is that "IT" is perceived as a free-standing topic in the school curriculum, rather than just as a background aid to learning. The hardness stems from the generality, hence complexity, of current computers, which in turn stems from the fact that most people only have at most one computer to use. This fact is set to change soon, as computers become pervasive commodities, not visibly, but hidden within specialised appliances. In this way, computers will become soft on their users, by just adding background value to everyday activities. The mobile telephone, digital television and games console are forerunners, and smart domestic and workplace appliances are imminent. Elimination of cabling, replaced by wireless communication between computers and by scavenging for electric power, will enhance the hidden nature of the computer, which will be able to collaborate invisibly with other computers, locally or globally.

For the computer architect, the terms "hard" and "soft" here have more physical meanings. Historically, the hardware of a computer has been the part that can be touched, usually electronic circuitry that implements a machine capable of obeying simple, general-purpose instructions. The software has been the intangible, flexible part, comprising particular programs of instructions to be obeyed by the hardware in order to carry out specific functions for a user's current needs. This traditional separation is now increasingly blurred, an effect catalysed in part by advances in silicon chip technology that allow not only a huge number of tiny communicating computational entities to co-exist on one chip, but a range of differing types of entity - electronic, mechanical, photonic. Decreasing hardness, that is, reduction in the amount of function fixed at the time of physical manufacture, means computers better equipped to be adaptive to their working environments, which includes their users and collaborating computers. So, perhaps unexpectedly, architects will design more flexible and general computers, but these will be presented to their users hidden within a range of less flexible and more specific appliances.

The talk covers the first hundred years of computing, beginning with the Bletchley Park codebreakers, the Manchester 'Baby', and the Cambridge Edsac. I look at developments over the next fifty years, bringing us up to date.

I then speculate on likely breakthroughs over the next few years, the next couple of decades, and finally some speculation on computing in 2050, the end of the first century of computing.

In psychological studies of cognition and behavior, variance is often considered a nuisance. It's just something that gets in the way of showing a statistically significant effect. However, population biologists recognize the fundamental importance of genetic variance for natural selection and evolution. By analogy to this biological perspective, it is easy to see that cognitive and behavioral variance can be good for cooperating groups of individuals. Using simple biologically appropriate neural networks, specific hypotheses can be shown more or less likely as the source of this variance. Indeed for a task requiring normal hippocampal function, a computational model of the hippocampus - with a particular form of randomization - improves performance, reproduces the population variance and, to surprising accuracy the frequency histogram of performance.

The last decades have seen the emergence of the "sea of objects" paradigm for structuring complex distributed systems on workstations and local area networks. In this approach, applications and system services are composed of and communicate among themselves through reliable and transparently accessible object interfaces, leading to the interaction of hundred or thousands of unstructured objects.

This approach has lead to major progress in software composability and reliability. Unfortunately, it is based on a number of assumptions that do not hold on wide-area networks. There, access to resources is intrinsically unreliable (because of failure, congestion, disconnected operation, etc.) and not transparent (because of variations in latency and bandwidth, hardware and software mobility, and the presence of firewalls). These characteristics amount to a new model of computation.

We discuss the challenges of computation on wide-area networks, and introduce a formalism, the Ambient Calculus, that matches some fundamental characteristics of wide-area networks and systems. Our approach (developed together with Andrew Gordon) reflects the intuition that to function satisfactorily on a wide-area network, the "sea of objects" must be partitioned and made hierarchical, internally mobile, and secure.

A large part of perception consists in inference. This is often done by Bayesian techniques, but for the purpose of connecting language and perception, it is useful to consider more abstract inference mechanisms, couched in logical terms.

A logic of perception is proposed here, as an extension of first order logic, which formalises an important feature of perception, viz. information reduction, or `filtering', by means of a new kind of generalised quantifier. We discuss applications of the proposed logic to Treisman's theory of `illusory conjunctions', the logic of perceptual expressions in natural language, and Herskovits' theory of prepositions

Understanding how to use computer graphics technology to portray 3D virtual humans has matured greatly in the past few years. Unlike the off-line, animator-intensive methods used in the special effects industry, real-time virtual humans are expected to exist and interact with us ``live.'' They can be our virtual selves (avatars), represent other people, or even function as autonomous helpers, teammates, or agents. We should be able to interact and communicate with them through modalities we already use, such as language. Such real-time virtual humans enable novel interactive educational and training applications. Various aspects and issues in real-time virtual humans will be discussed, such as gestural ``Effort'' control, autonomous attention, and coordinating individuals. Our current long range research is forging a deep connection between natural language instruction understanding and human action animation through a Parameterized Action Representation.

Approximate concepts can't have if-and-only-if definitions and usually don't even have definite extensions. Some approximate concepts can be refined by learning more and some by defining more and some by both. Approximate concepts are essential for representing common sense knowledge and doing common sense reasoning. Relations between approximate concepts and some of their refinements need to be represented. We use mathematical logic fortified with contexts as objects to represent facts involving approximate concepts. Further innovations in logic may be required to treat approximate concepts as flexibly in logic as people do in thought and language. 

A sentence involving an approximate concept may have a definite truth value even if the concept is ill-defined. For example, it is definite that Mount Everest was climbed in 1953 even though exactly what rock and ice is included in that mountain is ill-defined. Part of our goal is to discuss how solid intellectual structures can be built on swampy conceptual foundations. 

Theories using a concept of causality are usually approximate. 

Quantitative approximation one of the kinds considered---but not the most interesting or the kind that requires logical innovation. 

Fuzzy logic treats a particular kind of approximation.

Reasoning about the beliefs of other agents is thought to be an important part of just about any form of communicative or cooperative reasoning. But models of this reasoning have not been altogether successful. Philosophers (including Paul Grice and David Lewis) have recognized the importance of conclusions about others' beliefs, and psychologists (including Herb Clark) have illustrated the importance of mutual belief, but in general both groups have neglected to give any detailed account of how the reasoning works. Logicians have provided some puzzling theorems suggesting that under very general conditions, mutual belief seems to be unattainable.

I will suggest that part of the difficulty in modeling reasoning about the beliefs of other agents has to do with an inadequate model of single-agent beliefs. With a more modular model of single-agent attitudes, I believe it is possible to formalize a simulation model of reasoning about belief.

In the talk I will provide a general survey of the problem, as well as an overview of the solution that I have been developing over the last two years.

The expressiveness of higher-order, typed languages such as Haskell or ML makes them an attractive medium in which to write software components. Hitherto, however, their use has been limited by the all-or-nothing problem: it is hard to write just part of an application in these languages. 

Component-based programming using a binary standard such as Microsoft's Component Object Model (COM) offers a  solution to this dilemma, by specifying a language-independent interface between components. This paper reports about our experience with exploiting this opportunity in the purely-functional language Haskell. We describe a design for integrating COM components into Haskell programs, and we demonstrate why someone might want to script their COM components in this way.

The brain learns to convert the sensory input into internal representations of the causes of the input. It does this without having a teacher to specify what each internal neuron ought to be doing. I shall describe the "wake-sleep" algorithm which uses a generative model, embodied in top-down connections, to create target states for the internal neurons. These target states can then be used to train the bottom-up connections using a purely local learning rule. The generative model itself can also be learned using the same local rule. The algorithm can be interpreted as a way of minimizing the description length of the data.

As our society is developing ever more sophisticated information and knowledge technology, information and cognition are becoming pervasive integrative themes across the sciences and humanities. What is the essential structure of information (symbolic, graphic), and how do we communicate, learn, and transform it? A full under- standing of these phenomena would take us all the way from the physics of signal processing to computation in a wide variety of forms, and eventually, intelligent behavior in humans, machines, and social organizations. This is the grand challenge for a new discipline of Informatics, which seems to be slowly emerging.

With this grand panorama, critical foundational questions arise. We do not yet understand the notion of information, and there is not even a consensus on the basic questions that should be asked. I will discuss this issue of 'what should be our ambitions'. In particular, what exact theory can we have of the structure and flow of meaningful information? Logic and computer science in their present hold pieces of the puzzle (but: which puzzle?). I will discuss some examples from channel theory, dynamic logic, and game theory - and identify a broader agenda of challenges, broadening the scope of traditional semantics and algorithmic.