Larrie Hutton | Ohio University (original) (raw)

Papers by Larrie Hutton

A stochastic neural network, using the ordinary laws of thermodynamics, can develop a representat... more A stochastic neural network, using the ordinary laws of thermodynamics, can develop a representation of what might be called a state of cognitive equilibrium (Hutton, 1991). If it is assumed that there is a quanta] nature to behavior (and hence cognition) and that inputs are logarithmically transformed, three important and not obviously related psychological laws, usually treated as empirical generalizations, can be derived from first principles. Steven’s Law relates the subjective impression of a stimulus to its objective magnitude. Herrnstein’s equation relates absolute response rate to absolute reinforcement rate. The generalized matching law makes quantitative predictions about choice by assuming that a power law describes the relationship between behavioral states and relative value of alternatives. Assume a neural net with symmetric weights and a stochastic update rule. This results in relative output probabilities that are determined by a Boltzmann distribution. If inputs are logarithmically transformed, then the generalized matching law and Steven’s Law can be derived from the network equations. Assuming limited capacity, Herrnstein’s equation falls out. The results suggest heretofore unreported links between some apparently unrelated psychological laws, and show the importance of thermodynamic principles to states of cognitive/behavioral equilibria. They also show a plausible mechanism for the emergence of power law relationships that is not in any way restricted to treatments of cognitive and other behavioral states.

Johns Hopkins Apl Tech Dig, 1992

It is first pointed out that classical (statistical) techniques can be used to evaluate the perfo... more It is first pointed out that classical (statistical) techniques can be used to evaluate the performance of neutral-network (NN) classifiers. Second, it is noted that NN classifers often outperform classical techniques. It is also indicated that NN classifers may have advantages even when their ultimate performance on a training set can be shown to be no better than that of a classical classifer. Finally, it is suggested that methods which are routinely employed in statistics, but not in NN approaches, should also be adopted for the latter. 43 refs.

Neural network PC tools, Jan 3, 1990

Journal of the Experimental Analysis of Behavior, 1979

Previous research has shown that presenting response-independent positive reinforcers reduces the... more Previous research has shown that presenting response-independent positive reinforcers reduces the response rate of an operant maintained by positive reinforcement. The present experiment investigated a similar effect using shock-free time as a negative reinforcer. Brief shocks were delivered in the presence of a distinctive stimulus, and pigeon's key pecks were reinforced by the occasional presentation of a 2-minute shock-free period. Extra 2-minute shock-free periods were added independently of behavior. For each of three pigeons, response rate during shock-on periods declined with added shock-free periods; the more frequently the extra shock-free periods occurred the greater the decline in response rate. This outcome is predicted by extending the Law of Effect to include negative reinforcement.

Journal of the Experimental Analysis of Behavior, 1976

Rats were shocked at the rate of two per minute until they pressed a lever. In Experiment I, shoc... more Rats were shocked at the rate of two per minute until they pressed a lever. In Experiment I, shocks were delivered at variable-time intervals averaging 30 sec; in Experiment II, shocks were delivered at fixed-time intervals of 30 sec. A response produced an alternate condition for a fixed-time period. The shock frequency following a response, calculated over the whole alternate condition, was two per minute. The pattern of shocks in the alternate condition was controlled so that the first shock occurred at the same time as it would have occurred had the response not been emitted; the remaining shocks were delayed until near the end of the alternate condition. Bar pressing was acquired in both experiments. This finding is not explained by two-factor theories of avoidance and is inconsistent with the notion that overall shock-frequency reduction is necessary for negative reinforcement. The data imply that responding is determined by the integrated delays to each shock following a response versus the integrated delays to shock in the absence of a response.

Journal of the Experimental Analysis of Behavior, 1978

In the absence of responding, pigeons were shocked under a variable-time schedule. Responses on e... more In the absence of responding, pigeons were shocked under a variable-time schedule. Responses on either of two keys occasionally produced one minute of shock-free time. That is, pigeons' key pecks were reinforced with shock-free time under concurrent variable-interval schedules. The relative frequency of access to the one-minute shock-free periods was systematically manipulated. Pigeons tended to match both relative response rate and proportion of time spent on each key to the relative frequency of the shock-free periods. A bestfit linear regression equation accounted for over 95% of the variance in both relative response rate and time allocation. The data paralleled closely the results of concurrent schedules of positive reinforcement. These findings are consistent with a description of reinforcement as a transition to a higher-valued situation and suggest that common laws govern choice for both positive and negative reinforcement. Key words: avoidance, negative reinforcement, concurrent schedules, matching law, law of effect, key pecking, pigeons According to the matching law (Herrnstein, 1970) the rate of response controlled by a reinforcing alternative is equal to the proportion of reinforcers delivered for that alternative. This can be expressed as a relative rate by B1 __r1 (1) B2 r2 in which B1 and B2 represent the response rates associated with the two alternatives. The reinforcement value associated with alternative one and alternative two are represented by r, and r2, respectively. If r, and r2 are expressed as rate of reinforcement, Equation 1 predicts that relative response rate will be equal to the relative rate of reinforcement. Although Equation 1 is an empirical generalization based on the results of experiments involving positive reinforcement, a recent modification of the law of effect (Baum, 1973) applies also to experiments involving aversive stimuli. Baum defined reinforcement as a transition from a low-valued situation to a higher-"This research was supported in part by NIMH Grant No. lROI MH9593-1 and a grant (OURC 521

We investigated the effects of providing auxiliary knowledge (or “hints”), of varying the number ... more We investigated the effects of providing auxiliary knowledge (or “hints”), of varying the number of hidden nodes, and of providing secondary structure on the performance of feedforward networks with 0, 3, 6, and 9 hidden nodes. Data that permitted the prediction of diabetes from Pima Indian women served as inputs. By systematically adding secondary structure (not obtainable from the original data), we were able to show that convergence time was a function of the number of hidden nodes. The results suggest that neural nets learn “the easy things first” and that providing additional information may impair performance if secondary structure exists in the input data. We propose a model that is consistent with our results and that is also able to account for the common finding that performance on testing sets shows an initial increase followed by a gradual decline to an asymptote.

In ionospheric research, we must classify radar returns from the ionosphere as either suitable fo... more In ionospheric research, we must classify radar returns from the ionosphere as either suitable for further analysis or not. This time-consuming task has typically required human intervention. We tested several different feedforward neural networks to investigate the effects of network type (single-layer versus multilayer) and number of hidden nodes upon performance. As expected, the multilayer feedforward networks (MLFN'S) outperformed the single-layer networks, achieving 1000/0 accuracy on the training set and up to 980/0 accuracy on the testing set. Comparable figures for the single-layer networks were 94.50/0 and 92 0/0, respectively. When measures of sensitivity, specificity, and proportion of variance accounted for by the model are considered, the superiority of the MLFN'S over the single-layer networks is even more striking.

[1991] Computer-Based Medical Systems@m_Proceedings of the Fourth Annual IEEE Symposium

... RC Eberhart, R. W. Dobbins and L. V. Hutton The Johns Hopkins University Applied Physics Labo... more ... RC Eberhart, R. W. Dobbins and L. V. Hutton The Johns Hopkins University Applied Physics Laboratory Laurel, Maryland 20723 U. S. A. Abstract ... MePSS will consist of four main modules: diagnosis/treatment, record keeping, training, and an on-line medical library. ...

Neural Network PC Tools, 1990

Bulletin of the Psychonomic Society, 1976

A preformed artificial neural network shows a surprising similarity to anticipatory responses of ... more A preformed artificial neural network shows a surprising similarity to anticipatory responses of monkeys while tracking targets

Journal of The Experimental Analysis of Behavior, 1978

Journal of The Experimental Analysis of Behavior, 1976