Biomedical subjects

D J Volper

Publications and source records attributed to D J Volper.

4 recordsLinked to original sources

PubMed1987

Disjunctive models of Boolean category learning.

Four connectionistic/neural models which are capable of learning arbitrary Boolean functions are presented. Three are probably convergent, but of differing generalization power. The fourth is not necessarily convergent, but its empirical behavior is quite good. The time and space characteristics of the four models are compared over a diverse range of functions and testing conditions. These include the ability to learn specific instances, to effectively generalize, and to deal with irrelevant or redundant information. Trade-offs between time and space are demonstrated by the various approaches.

Brain↗

PubMed1987

Learning and using specific instances.

A biologically plausible method for rapidly learning specific instances is described. It is contrasted with a formal model of classical conditioning (Rescorla-Wagner learning/perception training), which is shown to be relatively good for learning generalizations, but correspondingly poor for learning specific instances. A number of behaviorally relevant applications of specific instance learning are considered. For category learning, various combinations of specific instance learning and generalization are described and analyzed. Two general approaches are considered: the simple inclusion of Specific Instance Detectors (SIDs) as additional features during perception training, and specialized treatment in which SID-based categorization takes precedence over generalization-based categorization. Using the first approach, analysis and empirical results demonstrate a potential problem in representing feature presence and absence in a symmetric fashion when the frequencies of feature presence and absence are very different. However, it is shown that by using the proper representation, the addition of SIDs can only improve the convergence rate of perceptron training, the greatest improvement being achieved when SIDs are preferentially allocated for peripheral positive and negative instances. Some further improvement is possible if SIDs are treated in a specialized manner.

Animals↗

PubMed1986

Linear function neurons: structure and training.

Three different representations for a thresholded linear equation are developed. For binary input they are shown to be representationally equivalent though their training characteristics differ. A training algorithm for linear equations is discussed. The similarities between its simplest mathematical representation (perceptron training), a formal model of animal learning (Rescorla-Wagner learning), and one mechanism of neural learning (Aplysia gill withdrawal) are pointed out. For d input features, perceptron training is shown to have a lower bound of 2d and an upper bound of dd adjusts. It is possible that the true upper bound is 4d, though this has not been proved. Average performance is shown to have a lower bound of 1.4d. Learning time is shown to increase linearly with the number of irrelevant or replicated features. The (X of N) function (a subset of linearly separable functions containing OR and AND) is shown to be learnable in d3 time. A method of utilizing conditional probability to accelerate learning is proposed. This reduces the observed growth rate from 4d to the theoretical minimum (for unmodified version) of 2d. A different version reduces the growth rate to about 1.7d. The linear effect of irrelevant features can also be eliminated. Whether such an approach can be made probably convergent is not known.

Animals↗

PubMed1986

Connectionistic models of Boolean category representation.

Several distinct connectionistic/neural representations capable of computing arbitrary Boolean functions are described and discussed in terms of possible tradeoffs between time, space, and expressive clarity. It is suggested that the ability of a threshold logic unit (TLU) to represent prototypical groupings has significant advantages for representing real world categories. Upper and lower bounds on the number of nodes needed for Boolean completeness are demonstrated. The necessary number of nodes is shown to increase exponentially with the number of input features, the exact rate of increase depending on the representation scheme. In addition, in non-recurrent networks, connection weights are shown to increase exponentially with a linear reduction in the number of nodes below approximately 2d. This result suggests that optimum memory efficiency may require unacceptable learning time. Finally, two possible extensions to deal with non-Boolean values are considered.

Artificial Intelligence↗