Search

Traditional methods for explaining programs provide explanations by converting to English the code of the program or traces of the execution of that code. While such methods can provide adequate explanations of what the program does or did, they typically cannot provide justifications of the code without resorting to canned-text explanations. That is, such systems cannot tell why what the system is doing is a reasonable thing to be doing.

In recent years many computer systems have been developed to assist in medical decision making. Two of these systems in particular, INTERNIST and the Present Illness Program (PIP), have been proposed as suitable for performing general medical diagnosis. However, there has been no way of comparing the performance of these two programs since the medical data used by the programs differs extensively.

We describe the motivation and overall organization of the OWL language for knowledge representation. OWL consists of a memory of concepts in terms of which all English phrases and all knowledge of an application domain are represented, a theory of English grammar which tells how to map English phrases into concepts, a parser to perform that mapping for individual sentences, and an interpreter to carry out procedures which are written in the same representational formalism. The system has been applied to the study of interactive dialogs, explanations of its own reasoning, and question answering.

The Internet has brought a new meaning to the term communities. Geography is no longer a barrier to international communications. However, the paradigm of meeting new interesting people remains entrenched in traditional means; meeting new interesting people on the Internet still relies on chance and contacts. This thesis explores a new approach towards matching users in online communities in an effective fashion. Instead of using the conventional feature vector scheme to profile users, each user is represented by a personalized concept hierarchy (or an ontology) that is learnt from the user's behavior in the system. Each concept hierarchy is then interpreted within the Information Theory framework as a probabilistic decision tree. The matching algorithm uses the Kullback-Leiber distance as a measure of deviation between two probabilistic decision trees. Thus, in an online community, where a personalized concept hierarchy represents each user, the Kullback-Leiber distance imposes a full- order rank on the level of similarity of all the users with respect to a particular user in question. The validity and utility of the proposed scheme of matching users is then applied in a set of simulations, using the feature-vector-overlap measure as a baseline. The results of the simulations show that the Kullback Leiber distance, when used in conjunction with the concept hierarchy, is more robust to noise and is able to make a stronger and more distinctive classification of users into similar groups in comparison to the conventional keyword-overlap scheme. A graphical agent system that relies upon the ontology-based interest matching algorithm, called the Collaborative Sanctioning Network, is also described in this thesis.

Qualitative analysis is the study of abstract causal reasoning. It explores the mechanisms whereby humans analyze complex systems abstractly, while ignoring unimportant and unknown low-level details. Previous research has focused on qualitative simulation techniques, analogous to numerical simulation, that use local information about a system to predict its short-term behavior. This thesis presents a new, calculus based, type of qualitative analysis, called qualitative mathematical reasoning. It derives functional descriptions of systems and uses them to predict global behavior.

We show how a computational system can be constructed to "reason," effectively and consequentially, about its own inferential processes. The analysis proceeds in two parts. First, we consider the general question of computational semantics, rejecting traditional approaches, and arguing that the declarative and procedural aspects of computational symbols (what they stand for, and what behaviour they engender) should be analysed independently, in order that they may be coherently related.

This paper proposes a welcome hypothesis: a computationally simple device is sufficient for processing natural language. Traditionally it has been argued that processing natural language syntax requires very powerful machinery. Many engineers have come to this rather grim conclusion: almost all working parsers are actually Turing Machines (TM). For example, Woods specifically designed his Augmented Transition Networks (ATNs) to be Turing Equivalent.

An ambitious, but intriguing, possibility for radically increasing the availability and adequacy of health case, while containing cost, is to use the computer as a consultant to augment and extend the skills of all health care providers. We propose to pursue a program of fundamental research in representation of knowledge, decision-making, problem-solving, program explanation and clinical cognition, to understand how to construct computer programs that, as an integral part of the health care system, can amplify the knowledge and reasoning powers of medical decision makers. We plan to apply the techniques so developed to problems in the diagnosis and therapy of acid/base and electrolyte disturbances, diagnosis of birth defects using an existing date-base of diseases and associated manifestations, the development of multi-modal cancer therapy protocols, and the application of the methods of decision analysis to produce general tools for physicians to use in analysing difficult clinical cases.

Current research in Artificial Intelligence has yielded computer programs which have potential to augment the physician's ability to diagnose illness. The medical diagnoses programs of the first generation contain medical facts representing associations between diseases and findings. A most important step is the development of computer programs that have models of physiological processes and have the ability to derive physiological justifications of observed signs and symptoms.

Clinical decision making involves a large, complex, and ever-changing body of knowledge. Characterizing such knowledge illuminates the representational and computational requirements for automated clinical decision analysis.

This memo is the text of a proposal from the MIT Laboratory for Computer Science Clinical Decision Making group to the National Library of Medicine, requesting support for a five-year program of research.

This thesis describes the English explanation facility of the OWL Digitalis Advisor, a program designed to advise physicians regarding digitalis therapy. The program is written in OWL, an English-based computer language being developed at MIT. The system can explain, in English, both the methods it uses and how those methods were applied during a particular session. In addition, the program can explain how it acquires information and tell the user how it deals with that information either in general or during a particular session.

For distributed systems at large and e-commerce systems in particular, ratings play an increasingly important role. Rating confer reputation measures about sources. This paper reports our formalization of the rating process. This paper argues that rating shuold be context- and individual- dependent quantities. In contrast to existing rating systems in many e-commerce or developer sites, our approach makes use of personalized and contextualized ratings for assessing source reputation. Our approach is based on a Bayesian probabilistic framework.

This paper describes a program, GENINFER, which uses belief networks to calculate risks of inheriting genetic disorders. GENINFER is based on Judea Pearl's [17] algorithm for fusion and propagation in probabilistic belief networks. These networks allow the effects of various pieces of information to be propagated and fused in such a way that, when equilibrium is reached, each proposition can be assigned a degree of believe consistent with the axioms of probability theory.

This report describes Paul, a computer text generation system designed to create cohesive text. The device used to a achieve this cohesion is lexical substitution. Through the use of syntactic and semantic information, the system is able to determine which type of lexical substitution will provide the necessary information to generate an understandable reference, while no providing so much information that the reference is confusing or unnatural.

Much of the medical knowledge in the first generation Al in Medicine programs is phenomenological; that is, it describes the associations among phenomena without knowledge of the underlying causal mechanisms. Although these AIM programs provide a good first approximation to the way clinicians reason, they fail to produce clinicians reasoning based on a deaper understanding of the phenomens.

Programs that make decisions need mechanisms for representing and reasoning about the desirability of the possible consequences of their choices. This work is an exploration of preference models based on utility theory. The framework presented is distinguished by a qualitative view of preferences and a knowledge-based approach to the application of utility theory. The design for a comprehensive preference modeler is implemented in part by the Utility Reasoning Package (URP), a collection of facilities for constructing and analyzing preference models.

BRAND X is a simple representation language implemented as a pure extension of LISP. BRAND X provides the following additional facilities over LISP: Unique and canonical structures, property lists for all objects, labels for all objects, and a syntax to express each of these, supported by a reader and printer. BRAND X is intended as an "assembly language" for representation languages, attempting to provide facilities generally found useful in the simplest manner, without any strong commitment to specific representational conventions.

Named-entity recognition (NER) aims at identifying entities of interest in a text. Artificial neural networks (ANNs) have recently been shown to outperform existing NER systems. However, ANNs remain challenging to use for non-expert users. In this paper, we present NeuroNER, an easy-to-use named-entity recognition tool based on ANNs. Users can annotate entities using a graphical web-based user interface (BRAT): the annotations are then used to train an ANN, which in turn predict entities' locations and categories in new texts. NeuroNER makes this annotation-training-prediction flow smooth and accessible to anyone.

Machine learning algorithms are often vulnerable to adversarial examples that have imperceptible alterations from the original counterparts but can fool the state-of-the-art models. It is helpful to evaluate or even improve the robustness of these models by exposing the maliciously crafted adversarial examples. In this paper, we present TextFooler, a simple but strong baseline to generate adversarial text. By applying it to two fundamental natural language tasks, text classification and textual entailment, we successfully attacked three target models, including the powerful pre-trained BERT, and the widely used convolutional and recurrent neural networks. We demonstrate three advantages of this framework: (1) effective—it outperforms previous attacks by success rate and perturbation rate, (2) utility-preserving—it preserves semantic content, grammaticality, and correct types classified by humans, and (3) efficient—it generates adversarial text with computational complexity linear to the text length.1

The increasing availability of large collections of electronic health record (EHR) data and unprecedented technical advances in deep learning (DL) have sparked a surge of research interest in developing DL based clinical decision support systems for diagnosis, prognosis, and treatment. Despite the recognition of the value of deep learning in healthcare, impediments to further adoption in real healthcare settings remain due to the black-box nature of DL. Therefore, there is an emerging need for interpretable DL, which allows end users to evaluate the model decision making to know whether to accept or reject predictions and recommendations before an action is taken. In this review, we focus on the interpretability of the DL models in healthcare. We start by introducing the methods for interpretability in depth and comprehensively as a methodological reference for future researchers or clinical practitioners in this field. Besides the methods' details, we also include a discussion of advantages and disadvantages of these methods and which scenarios each of them is suitable for, so that interested readers can know how to compare and choose among them for use. Moreover, we discuss how these methods, originally developed for solving general-domain problems, have been adapted and applied to healthcare problems and how they can help physicians better understand these data-driven technologies. Overall, we hope this survey can help researchers and practitioners in both artificial intelligence and clinical fields understand what methods we have for enhancing the interpretability of their DL models and choose the optimal one accordingly. This article is categorized under: Cancer > Computational Models.