B. Chandrasekaran

Generic tasks for knowledge-based reasoning: the “right” level of abstraction for knowledge acquisition

Our research strategy has been to identify generic tasks—basic combinations of knowledge structures and inference strategies that are powerful for solving certain kinds of problems. Our strategy is best understood by considering the “interaction problem”, that representing knowledge for the purpose of solving some problem is strongly affected by the nature of the problem and by the inference strategy to be applied to the knowledge. The interaction problem implies that different knowledge-acquisition methodologies will be required for different kinds of reasoning, e.g. a different knowledge-acquisition methodology for each generic task. We illustrate this using the generic task of hierarchical classification. Our proposal and the interaction problem call into question many generally held beliefs about expert systems such as the belief that the knowledge base should be separated from the inference engine.

Deep versus compiled knowledge approaches to diagnostic problem-solving

Most of the current generation expert systems use knowledge which does not represent a deep understanding of the domain, but is instead a collection of “pattern?action” rules, which correspond to the problem-solving heuristics of the expert in the domain. There has thus been some debate in the field about the need for and role of “deep” knowledge in the design of expert systems. It is often argued that this underlying deep knowledge will enable an expert system to solve hard problems. In this paper we consider diagnostic expert systems and argue that given a body of underlying knowledge that is relevant to diagnostic reasoning in a medical domain, it is possible to create a diagnostic problem-solving structure which has all the aspects of the underlying knowledge needed for diagnostic reasoning “compiled” into it. It is argued this compiled structure can solve all the diagnostic problems in its scope efficiently, without any need to access the underlying structures. We illustrate such a diagnostic structure by reference to our medical system MDX. We also analyze the use of these knowledge structures in providing explanations of diagnostic reasoning.

On dimensionality and sample size in statistical pattern classification

Abstract The basic question of how to optimally make use of a finite number of available samples in designing pattern recognition systems is considered. This has several components: optimal use of the samples for design and testing; and the relationship between the number of measurements and the number of samples for various probability structural constraints. A spectrum of possibilities has been demonstrated, placing several apparently conflicting recent results in perspective.

Conceptual Representation of Medical Knowledge for Diagnosis by Computer: MDX and Related Systems

Publisher Summary This chapter describes an approach to the design of medical decision-making systems based on the notion of conceptual structures for knowledge representation. The chapter provides an overview, from a theoretical viewpoint, of the conceptual structure methodology and describes the functioning of the systems that have been developing to give concreteness to the theoretical ideas. The central system in this group of systems is called MDX, which is a diagnostic system, that is, it attempts to classify a given case as an element of a disease taxonomy. This system interacts with two other systems during its problem solving, PATREC and RADEX, the former a knowledge-based patient database system that answers MDXs queries about patient data, and the latter a radiological consultant which helps MDX in the interpretation of various kinds of imaging data. Both PATREC and RADEX are invoked by MDX as needed, but MDX is in control of the overall diagnostic process. The chapter discusses the inadequacies of medical reasoning approaches based on Bayesian approaches.

Representing function: Relating functional representation and functional modeling research streams

This paper is an informal description of some recent insights about what a device function is, how it arises in response to needs, and how function arises from the structure of a device and the functions of its components. These results formalize and clarify a set of contending intuitions about function that researchers have had. The paper relates the approaches, results, and goals of this stream of research, called functional representation (FR), with the functional modeling (FM) stream in engineering. Despite the occurrence of the term function in the two streams, often the results and techniques in the two streams appear not to have much to do with each other. I argue that, in fact, the two streams are performing research that is mutually complementary. FR research provides the basic layer for device ontology in a formal framework that helps to clarify the meanings of terms such as function and structure, and also to support representation of device knowledge for automated reasoning. FM research provides another layer in device ontology, by attempting to identify behavior primitives that are applicable to subsets of devices, with the hope that functions can be described in those domains with an economy of terms. This can lead to useful catalogs of functions and devices in specific areas of engineering. With increased attention to formalization, the work in FM can provide domain-specific terms for FR research in knowledge representation and automated reasoning.

Microbenchmark performance comparison of high-speed cluster interconnects

Todays distributed and high-performance applications require high computational power and high communication performance. Recently, the computational power of commodity PCs has doubled about every 18 months. At the same time, network interconnects that provide very low latency and very high bandwidth are also emerging. This is a promising trend in building high-performance computing environments by clustering - combining the computational power of commodity PCs with the communication performance of high-speed network interconnects. There are several network interconnects that provide low latency and high bandwidth. Traditionally, researchers have used simple microbenchmarks, such as latency and bandwidth tests, to characterize a network interconnects communication performance. Later, they proposed more sophisticated models such as LogP. However, these tests and models focus on general parallel computing systems and do not address many features present in these emerging commercial interconnects. Another way to evaluate different network interconnects is to use real-world applications. However, real applications usually run on top of a middleware layer such as the message passing interface (MPI). Our results show that to gain more insight into the performance characteristics of these interconnects, it is important to go beyond simple tests such as those for latency and bandwidth. In future, we plan to expand our microbenchmark suite to include more tests and more interconnects.

Representation, organization, and use of topographic models of physical spaces for route planning

ROUTER 1 is a new route-planning system. The focus is on knowledge and its processing-more specifically, on the representation and organization of topographic models of physical Spaces and their use for route planning. The route-planning task is decomposed into the subtasks of direction-finding and route-finding. In direction-finding, ROUTER 1 finds the neighborhoods of the initial and goal locations and the direction of the goal location relative to the initial location. In route-finding, ROUTER 1 uses knowledge of the relative direction as a heuristic for selecting pathways. It first selects a high-level pathway from the neighborhood of the initial location to the neighborhood of the goal location, and, then, progressively adds more details to the growing route until a complete legal route is synthesized. This organizational scheme and reasoning method enables ROUTER 1 to solve nontrivial route-planning tasks efficiently and effectively.<<ETX>>

Micro-benchmark level performance comparison of high-speed cluster interconnects

In this paper we present a comprehensive performance evaluation of three high speed cluster interconnects: Infini-Band, Myrinet and Quadrics. We propose a set of micro-benchmarks to characterize different performance aspects of these interconnects. Our micro-benchmark suite includes not only traditional tests and performance parameters, but also those specifically tailored to the interconnects advanced features such as user-level access for performing communication and remote direct memory access. In order to explore the full communication capability of the interconnects, we have implemented the micro-benchmark suite at the low level messaging layer provided by each interconnect. Our performance results show that all three interconnects achieve low latency, high bandwidth and low host overhead. However, they show quite different performance behaviors when handling completion notification, unbalanced communication patterns and different communication buffer reuse patterns.

A heuristic strategy for developing human facial images on a CRT

Abstract Sketching a recognizable human face involves artistic talents and an intuitive knowledge of which aspects of the face are important in recognition. A man-machine system, called WHATSISFACE, has been developed with which man-machine system, called WHATSISFACE, has been developed with which a nonartist can create, on a graphic display, any male Caucasian facial image resembling the face of a photograph in front of him. The computer system contains pre-stored facial features, an average face used as a starting point and a heuristic strategy which guides the user through a carefully constructed sequence of questions, choices and feature manipulations. The user makes all the visual decisions and can change the individual features or hierarchically organized sets of features using analog input devices.

MIBA: A Micro-Benchmark Suite for Evaluating InfiniBand Architecture Implementations

Recently, InfiniBand Architecture (IBA) has been proposed as the next generation interconnect for I/O and inter-process communication. The main idea behind this industry standard is to use a scalable switched fabric to design the next generation clusters and servers with high performance and scalability. This architecture provides various types of new mechanisms and services (such as multiple transport services, RDMA and atomic operations, multicast support, service levels, and virtual channels). These services are provided by components (such as queue pairs, completion queue, and virtual-to-physical address translations) and their attributes. Different implementation choices of IBA may lead to different design strategies for efficient implementation of higher level communication layer/libraries (such as Message Passing Interface (MPI), sockets, and distributed shared memory). It also has an impact on the performance of applications.