Are existing knowledge discovery systems destined to suffer the Cassandra effect - where the system accurately predicts the future but is unable to change it? Despite the sporadic success of automated discovery systems, few studies have systematically explored the socio-technical environments in which a discovery tool will ultimately be embedded. Modeling the day-to-day activities of experienced scientists as they develop and verify hypotheses provides both a glimpse into the human reasoning processes that surround discovery and a deeper understanding of the characteristics that are required for a discovery system to be successful. In this talk, I will describe how experienced faculty in chemistry and chemical engineering characterize discovery, how they arrive at their research questions, and the processes they use to transform an initial idea into a subsequent publication. By using a published article as the unit of analysis we were able to explore the creative processes that come into play when scientists engage in what Kuhn would call "normal" science. I will conclude the talk by describing the features of a discovery system that are critical to support the creative endeavors of scientists.
Models link theoretical principles to empirical evidence, but how can one build rich explanatory structures from a collection of laws, processes, entities, and so on? In prior work on inductive process modeling, my colleagues and I put forth an approach founded upon eliminative induction, which considers all structures relevant to the modeled scenario. This approach, although effective, leads to explanations that scientists reject regardless of their accuracy. Probing more deeply, we found dependencies among the processes and the features of the target scenario that strongly constrain the space of plausible structures. In this talk, I will introduce these structural constraints as an explicit form of scientific knowledge, show how one can discover new constraints using well tested computational methods, and suggest how these constraints affect scientific creativity.
In the the Combined Reasoning Group (www.doc.ic.ac.uk/crg), we combine theorem provers, machine learning systems, constraint solvers, SAT solvers, model generators and computer algebra packages so that the whole is more than the sum of the parts. I will survey a number of projects we have undertaken recently with such combinations, including applications to AI techniques (non-theorem proving, CSP reformulation and automatic invention of fitness functions), and applications to discovery in pure mathematics (in particular classification of finite algebras). I will describe some lessons learned from integrating these systems for creative tasks, and discuss why we can expect more creativity from combined rather than stand-alone AI systems.
One measure of a good conjecture or problem in mathematics is that it lead to new tools and techniques for making advances in the area under investigation. Often these tools and techniques have the reciprocating effect of inspiring new problems and conjectures. In this talk, after describing some types of graph theory problems and conjectures that are of current interest, I will describe how one computer program, Graffit.pc, evaluates and poses two types of graph theoretical conjectures. As time permits I will propose enhancements for Graffiti.pc that could be helpful for further evaluating and working on its conjectures.
Imagination is more important than knowledge — Albert Einstein
Imagination appears to be a fundamental process of scientific creativity. When scientists build new theories they often view familiar objects, relations and processes in new and imaginative ways, for example, viewing the earth as a point mass and imagining the concept of anti-matter. TORQUE is a computer program that emulates verbal protocols of physicists addressing difficult problems pertaining to spring systems. In the course of its problem solving, TORQUE imagines new forms of springs. In this talk, I will describe how TORQUE uses its mental models, working memory, analogical reasoning, and control of processing as building blocks for productive imagination. I will also discuss the insights TORQUE provides into the processes of imagination in scientific creativity and their implications for designing computational environments for aiding scientific creativity.
Scientists and engineers are increasingly involved in large, collaborative projects that link multiple laboratories and expertises, and also often involve stakeholders outside of the technological communities. The development of convergent technologies to enhance human performance is an example. Such collaborations face Kuhn's age-old 'problem of incommensurability', or the difficulty in communicating across research paradigms.
In addition to the growing literature in cognitive science on scientific and technological thinking (M. E. Gorman, 2006), there is a new research program in science and technology studies (STS) devoted to Studies of Expertise and Experience (SEE). This presentation links these two literatures to derive a framework for future research (M. E. Gorman, 2008). Different expertise can surmount apparent incommensurabilities by forming trading zones around the development of new technological systems like radar, particle detectors and astronomical instruments.
This presentation will present a preliminary taxonomy of such trading zones and show trajectories by which one type can morph into another. The study of such zones, and of other collaborative networks, could be facilitated by computational models, and their effectiveness will be enhanced by collaborative IT.
Abduction and induction are two forms of reasoning that, along with deduction, are closely associated with the business of scientific discovery. To a first approximation, abduction formalises the task of reasoning from observed effects to possible explanations, while induction formalises the task of reasoning from known samples to probable generalisations. After reviewing the roles of abductive and inductive inference in the philosophy of science, this talk will focus on their realisation in the setting of computational logic. A recent technique for integrating abductive and inductive inference within a common nonmonotonic logic programming framework will be described and illustrated on some biologically motivated discovery problems.
Computational approaches to scientific creativity have built upon wonderful interdisciplinary work with the psychology of science and philosophy of science. Much less of the thinking about models of science has been built upon the psychology of design or computational models of design, even though the same people doing relevant work on science have also done relevant work on design. I will review recent work on psychological studies and computational models of design, with a particular focus on creativity in engineering design, and draw out a number of suggestions for how work in computational models of science could be advanced by this kind of analogical reasoning (e.g., relating to interactions of fixation, analogy, and the surrounding environment).
When applied to relevant tasks, knowledge bases associated with complex, scientific domains are often shown to be incomplete, which requires one to acquire new knowledge, and inconsist, which necessitates the refinement of the existing knowledge. Assuming that the topic is at the domain's cutting edge, it is generally only possible to provide a partial domain theory, so one must involve an oracle (i.e., a domain expert) to select from alternative refinements and to provide the missing knowledge. For these reasons, we refer to these systems as Cooperative Knowledge Acquisition and Knowledge Refinement Systems.
In this talk I will give an overview of the systems which we, at Aberdeen, have built over the last decade, which can acquire and refine knowledge bases expressed as "classical-rules", cases, taxonomies and qualitative graphs. I will then discuss the use of such systems/techniques as:
The one-node search is an open-ended text mining technique in which an investigator begins with a single set of articles A representing a scientific problem, and seeks to find information published in some disparate field C (unknown to the investigator at the outset) that suggests a novel hypothesis that contributes to solving the problem. Most existing strategies extract features from A (these are so-called B-terms, which may be e.g., title terms, indexing terms, or topical themes) that are used to query a digital library to identify a set of Bi-literatures that, in turn, has features extracted from it to identify candidate Ci-literatures. Most proposed one-node strategies use B-terms for ranking Ci's as well, to find the ones that are most similar to A in some sense. The user (a human) then examines articles in Ci and A to see if they contain implicit assertions that would be readily recognized as a prediction (i.e., a possible discovery) worthy of testing in the laboratory. Such systems only partially emulate the thought processes and priorities of working scientists. I will discuss a different strategy that does not employ B-terms to identify candidate Ci literatures, and that defines multiple "interestingness" measures on candidate Ci literatures which allows their ranking in a manner not solely determined by their relation to B-terms. This allows more flexibly for different models of scientific hypothesis formulation, by weighting and combining "interestingness" measures in different ways.
Human creativity is the result of neural processes that can be investigated by developing computational models. Theoretical neuroscience is exploring novel ideas about representation and computation that have the potential to illuminate the process of discovery. Neural representations can be multimodal, integrating verbal information with sensory, motor, and emotional encodings. I will describe a new neurocomputational model of hypothesis generation that suggests how large populations of neurons might be creative.
In this talk, I will discuss the task of establishing a mathematical model of an observed system or phenomena. As opposed to traditional creative tasks or domains, such as poetry, painting, music, or pure mathematical reasoning, mathematical modeling has a rigorous definition of valid artifacts. The methods for automated modeling that I will present integrate these validity constraints so that they consider only those mathematical models that are well formed, valid, and plausible in the domain of interest. I will discuss how and whether the strictness of computational approaches to automated modeling leaves room for creativity.
One of the most fundamental areas of creativity in science is the scientist's ability to use their imagination. We explore imagination by examining how experts and novices in different scientific and engineering domains create and manipulate spatial representations while using scientific visualizations. We also describe a computational system that facilitates some types of imagination and show the advantages and disadvantages of our (and perhaps other) systems.
with Susan Trickett and Christian Schunn.
It can be argued that computational modeling of the creative process in science can be facilitated by an analysis of the cognitive processes underlying scientific advances. We use the term discovery to describe the process of creativity in science; e.g., Watson and Crick discovered the double helix. When describing the arts, in contrast, we use the term create; e.g., Picasso created Guernica, his great anti-war painting of 1937. The first part of my presentation will examine the question of whether there is a dichotomy between artistic and scientific creativity, as implied by those different linguistic constructions, or might we say that Watson and Crick created the double helix? I will then discuss Watson and Crick's achievement, to demonstrate how one can use case studies of seminal creative advances to further our understanding of the creative process, with implications for modeling the creative process in science (and, perhaps, in the arts).
Back to Symposium Home Page