Though the experiences of life exhibit unceasing variety, people are able to find constancy and deal with their world in a regular and predictable manner. This thesis promotes the hypothesis that the necessary abstractions can be learned. The specific task studied is inducing a concept description from examples. A model is presented. that relies on a weighted, symbolic description of concepts. Though the description is distributed, novel examples are classified holistically by combining each portion's contribution. Each new example also refines the concept description: internal weights are updated and new symbolic structures are introduced. These actions improve description quality as measured by classification accuracy over novel examples.

Initially the concept description is highly distributed, being composed of many simple components. As learning progresses, sophisticated descriptive structures are added, and eventually the description is coalesced into a few highly predictive components. This qualitative change in the representation of the concept is a unique feature of the model.

The model extends previous work by allowing for noisy examples, unknown values, and concept change over time. To bolster claims of robustness, several experiments illustrating the model's behavior are reported. Key results illustrate that the model should scale-up to larger tasks than those studied and have a number of potential applications.

Determining the degree to which two events are interrelated is a common subtask for artificial intelligence systems, especially learning systems. This note examines four correlational measures which allow quantization of the relationships between events. Despite the fact that the measures have diverse motivations and formulations, they all indicate irrelevance precisely at the point of statistical independence.

Determining which features in an environment are salient given a task, salience assignment, is a central problem in Machine Learning. A related phenomenon, contingency (the conditions under which relative salience among environmental features is acquired), is central to learning and memory in animal psychology. This paper presents an analysis of a set of empirical data on contingency and an algorithm for the salience assignment problem. The algorithm presented is implemented in a working computer program which interacts with a simulated environment to produce contingent associative learning corresponding to relevant behavioral data. The model also makes specific empirical predictions that can be experimentally tested.

Incremental learning from examples in a noisy domain is a difficult problem in Machine Learning. In this paper we divide the task into two subproblems and present a combination of numeric and symbolic approaches that yields robust learning of boolean characterizations. Our method has been implemented in a computer program, and we plot its empirical learning performance in the presence of varying amounts of noise.

Contingency (the learned relative salience of environmental features) and latency (the learned timing of response to stimuli) are central phenomena of learning and memory. This paper provides a computational analysis of, and algorithms for, a set of empirical data on contingency and latency in classical and instrumental conditioning. These analyses are presented within the framework of an information-processing architecture that describes a set of modules which operate in parallel and asynchronously to store, retrieve and modify experiential information. The architecture (called 'CEL', for 'Components of Experiential Learning') provides a way of making explicit the interactions among a number of otherwise separate algorithms for related phenomena. The modules comprising the architecture each emerge from the operation of an indexed network memory. The algorithms presented are also implemented in working computer programs that interact with a simulated environment to produce contingent associative learning and differential response latencies that correspond to the relevant behavioral data. The model makes a number of specific empirical predictions that can be experimentally tested.

The problem of deciding which of many possible features in a training sequence are the salient or predictive ones is a well-known problem in machine learning. This problem of salience assignment is difficult when attempting to learn in an unpredictable and reactive environment. Components of Experiential Learning (CEL) is a framework for the development of computational theories of learning in this type of environment [Granger 1983, Granger and McNulty 1984). In this paper, we review the CEL processes and explain a specific computer model LURN (Learning by Unconscious ReasoNing) which illustrates the use of an incremental method for performing salience assignment. An additional con-straint on salience assignment arises from experimental results in animal psychology which indicate that living learning engines (i.e., humans and animals) make a sharp distinction between simple pairing (or strengthening) of associated events versus contingent salience assignment [Rescorla 1966). LURN calculates the contingent predictiveness of features in a noisy environment, in accordance with these experimental results.