Joachim Diederich

Survey and critique of techniques for extracting rules from trained artificial neural networks

It is becoming increasingly apparent that, without some form of explanation capability, the full potential of trained artificial neural networks (ANNs) may not be realised. This survey gives an overview of techniques developed to redress this situation. Specifically, the survey focuses on mechanisms, procedures, and algorithms designed to insert knowledge into ANNs (knowledge initialisation), extract rules from trained ANNs (rule extraction), and utilise ANNs to refine existing rule bases (rule refinement). The survey also introduces a new taxonomy for classifying the various techniques, discusses their modus operandi, and delineates criteria for evaluating their efficacy.

The truth will come to light: directions and challenges in extracting the knowledge embedded within trained artificial neural networks

To date, the preponderance of techniques for eliciting the knowledge embedded in trained artificial neural networks (ANNs) has focused primarily on extracting rule-based explanations from feedforward ANNs. The ADT taxonomy for categorizing such techniques was proposed in 1995 to provide a basis for the systematic comparison of the different approaches. This paper shows that not only is this taxonomy applicable to a cross section of current techniques for extracting rules from trained feedforward ANNs but also how the taxonomy can be adapted and extended to embrace a broader range of ANN types (e.g., recurrent neural networks) and explanation structures. In addition the paper identifies some of the key research questions in extracting the knowledge embedded within ANNs including the need for the formulation of a consistent theoretical basis for what has been, until recently, a disparate collection of empirical results.

Authorship Attribution with Support Vector Machines

In this paper we explore the use of text-mining methods for the identification of the author of a text. We apply the support vector machine (SVM) to this problem, as it is able to cope with half a million of inputs it requires no feature selection and can process the frequency vector of all words of a text. We performed a number of experiments with texts from a German newspaper. With nearly perfect reliability the SVM was able to reject other authors and detected the target author in 60–80% of the cases. In a second experiment, we ignored nouns, verbs and adjectives and replaced them by grammatical tags and bigrams. This resulted in slightly reduced performance. Author detection with SVMs on full word forms was remarkably robust even if the author wrote about different topics.

KRITON: a knowledge-acquisition tool for expert systems

A hybrid system for automatic knowledge acquisition for expert systems is presented. The system integrates artificial intelligence and cognitive science methods to construct knowledge bases employing different knowledge representation formalisms. For the elicitation of human declarative knowledge, the tool contains automated interview methods. The acquisition of human procedural knowledge is achieved by protocol analysis techniques. Textbook knowledge is captured by incremental text analysis. The goal structure of the knowledge elicitation methods is an intermediate knowledge-representation language on which frame, rule and constraint generators operate to build up the final knowledge bases. The intermediate knowledge representation level regulates and restricts the employment of the knowledge elicitation methods. Incomplete knowledge is laid open by patterndirected invocation methods (the intermediate knowledge base watcher) triggering the elicitation methods to supplement the necessary knowledge.

Explanation and artificial neural networks

Abstract Explanation is an important function in symbolic artificial intelligence (AI). For instance, explanation is used in machine learning, in case-based reasoning and, most important, the explanation of the results of a reasoning process to a user must be a component of any inference system. Experience with expert systems has shown that the ability to generate explanations is absolutely crucial for the user acceptance of Al systems. In contrast to symbolic systems, neural networks have no explicit, declarative knowledge representation and therefore have considerable difficulties in generating explanation structures. In neural networks, knowledge is encoded in numeric parameters (weight) and distributed all over the system. It is the intention of this paper to discuss the ability of neural networks to generate explanations. It will be shown that connectionist systems benefit from the explicit coding of relations and the use of highly structured networks in order to allow explanation and explanation components (ECs). Connectionist semantic networks (CSNs), i.e. connectionist systems with an explicit conceptual hierarchy, belong to a class of artificial neural networks which can be extended by an explanation component which gives meaningful responses to a limited class of “How” questions. An explanation component of this kind is described in detail.

Rule Extraction from Support Vector Machines

Rule Extraction from Support Vector Machines: An Introduction.- Rule Extraction from Support Vector Machines: An Overview of Issues and Application in Credit Scoring.- Algorithms and Techniques.- Rule Extraction for Transfer Learning.- Rule Extraction from Linear Support Vector Machines via Mathematical Programming.- Rule Extraction Based on Support and Prototype Vectors.- SVMT-Rule: Association Rule Mining Over SVM Classification Trees.- Prototype Rules from SVM.- Applications.- Prediction of First-Day Returns of Initial Public Offering in the US Stock Market Using Rule Extraction from Support Vector Machines.- Accent in Speech Samples: Support Vector Machines for Classification and Rule Extraction.- Rule Extraction from SVM for Protein Structure Prediction.

Lessons from Past, Current Issues, and Future Research Directions in Extracting the Knowledge Embedded in Artificial Neural Networks

Active research into processes and techniques for extracting the knowledge embedded within trained artificial neural networks has continued unabated for almost ten years. Given the considerable effort invested to date, what progress has been made? What lessons have been learned? What direction should the field take from here? This paper seeks to answer these questions. The focus is primarily on techniques for extracting rule-based explanations from feed-forward ANNs since, to date, the preponderance of the effort has been expended in this arena. However the paper also briefly reviews the broadening overall agenda for ANN knowledge-elicitation. Finally the paper identifies some of the key research questions including the search for criteria for deciding in which problem domains these techniques are likely to out-perform techniques such as Inductive Decision Trees.

Knowledge-based knowledge elicitation

A method for using the advantages of domain-specific knowledge acquisition for a general purpose knowledge acquisition tool is introduced. To adapt the knowledge acquisition tool for a specific application and a specific problem solving strategy (e.g. heuristic classification, such diagnostic strategies as establish and refine), acquisition knowledge bases (AKBs) are integrated in the system to guide the employment of different knowledge elicitation methods (interview techniques, protocol analysis, semantic text analysis and learning mechanisms). Acquisition knowledge bases are predefined deep models, consisting of structured objects to represent important concepts of a domain. These knowledge bases are used in addition to the already acquired knowledge to trigger specific elicitation methods by an analysis of incompleteness and inconsistency of the existing knowledge in the system. Furthermore, methods for integrating these kinds of knowledge acquisition tools with machine learning approaches are discussed.

Accent Classification Using Support Vector Machines

Accent is the pattern of pronunciation and acoustic features in speech which can identify a persons linguistic, social or cultural background. It is an important source of inter-speaker variability, and a particular problem for automated speech recognition. Current approaches to the identification of speaker accent may require specialised linguistic knowledge or analysis of the particular speech contrasts, and often extensive pre-processing on large amounts of data. An accent classification system using time-based segments consisting of Mel Frequency Cepstral Coefficients as features and employing Support Vector Machines is studied for a small corpus of two accents of English. On one- to four-second audio samples from three topics, accuracy in the binary classification task is up to 75% to 97.5%, with very high recall and precision. Its use with mis-matched content is at best 85% with a tendency towards majority-class classification if the accent groups are significantly imbalanced.

Knowledge extraction and recurrent neural networks: an analysis of an Elman network trained on a natural language learning task

We present results of experiments with Elman recurrent neural networks (Elman, 1990) trained on a natural language processing task. The task was to learn sequences of word categories in a text derived from a primary school reader. The grammar induced by the network was made explicit by cluster analysis which revealed both the representations formed during learning and enabled the construction of state-transition diagrams representing the grammar. A network initialised with weights based on a prior knowledge of the texts statistics, learned slightly faster than the original network.