Richard B. Millward

A physical system approach to recognition memory for spatially transformed faces

Abstract Recognition memory for familiar and unfamiliar spatial frequency filtered faces was examined using a combination of experiments and computer simulations based on the physical-systems approach to memory. The results of the experiments showed a significant interaction in recognition performance between the frequency characteristics of the learning and test pictures of the faces. In general, recognition transfer between normal and lowpass faces was good. Transfer between normal and high-pass faces was not as good but improved with the familiarity of the faces. Transfer between high-pass and low-pass faces was poor. A simple linear associative model of the experiment which used only basic visual information about the faces as input produced a pattern of results similar to those seen in the experiments. Memory enhancers aimed at improving the encoding and storage mechanisms of the model produced results similar to those seen when observers were familiar with the faces. These results are discussed in terms of the literature on recognition of faces which have been transformed in different ways and in terms of some psychological interpretations of codes and algorithms used in connectionist models.

Latency in a modified paired-associate learning experiment

Summary Latencies were analyzed in a 12-item 2-response paired-associate experiment with 20 S s. The basic assumption of the one-element model that associations occur in an all-or-none fashion suggests that the items be aligned according to the trial of the last error and then averaged for each trial. The latencies before the last error were large and constant but declined exponentially after the last error. These results seem to support an all-or-none interpretation of learning. The data were in good agreement with a number of statistics for the one-element model, but not the linear model. However, the proportion of errors before the last error, the so-called stationarity curve, was not constant as required by the one-element model, nor was it adequately described by the simple linear model.

PROBABILITY LEARNING: CONTINGENT-EVENT SCHEDULES WITH LAGS

The importance of memory for the immediately preceding sequence of events has been a recurrent theme in the two-choice prediction experiment. The degree to which subjects attend to preceding events, the manner in which they process the information contained in the event sequence, and the role that the past events play in determining prediction responses are all issues of considerable theoretical importance. There have been three basic ways to investigate the dependency between preceding events and prediction responses. The simplest and most indirect is statistical, as in the analyses of event patterns and recency curves (changes in response proportions as a function of the length

The direct measurement of hypothesis-sampling strategies ☆

Abstract The concept-identification (CI) literature supports a hypothesis-sampling theory. Hypotheses based on attributes, sample sizes greater than one, and nonreplacement of eliminated hypotheses all occur. An experimental procedure was developed to measure hypothesis-sampling directly where the subject is allowed to select attributes that he wishes to see, then randomly selected values on the selected attributes are presented. Generally, the average sample size before the trial of the last error (TLE) does not change but does change after TLE. The probabilities of eliminating inconsistent hypotheses and keeping consistent hypotheses increase over problems. The proportion of eliminated attributes which are resampled decreases over problems. Individual subjects are extremely varied in sample size, efficiency measures, and resampling tendencies. The rate of solution, TLE, is related to the efficiency and resampling measures but not to the sample size.

Attribute elimination strategies for concept identification with practiced subjects

Abstract The experiment reported in this paper investigated the strategies employed by well-practiced subjects in solving simple concept-identification (CI) problems. Fortyseven high school subjects were run for 10 hours on repeated 12-dimensional CI problems. Performance improved for the first 20 problems and remained stable there-after. The presolution response probability was constant, indicating an all-or-none solution process. A model based on the elimination of attributes, or dimensions, is developed to describe the asymptotic CI behavior. This model postulates that the subject tests samples of s attributes simultaneously, and eliminates from consideration those attributes that are inconsistent with the observed classification pattern. This elimination occurs regardless of the correctness of the overt response. Attributes thus eliminated are not immediately resampled for test, but remain excluded from consideration for an average of the next l samples. The selection of a sample containing the relevant attribute leads to solution of the problem. The range of individual subject solution rates was sufficiently large that no single choice of s and l could fit all subjects. The large number of problems available for a single subject, however, allows parameter pairs to roughly be assigned to subjects. The model is also related to some general theories of human information processing and memory.

Derivations of learning statistics from absorbing Markov chains

Learning-process statistics for absorbing Markov-chain models are developed by using matrix methods exclusively. The paper extends earlier work by Bernbach by deriving the distribution of the total number of errors, u-tuples, autocorrelation of errors, sequential statistics, and the expectation and variance of all statistics presented. The technique is then extended to latency derivations including the latencies of sequential statistics. Suggestions are made for using the sequential-statistic algorithm in a maximum-likelihood estimation procedure. The technique is important because statistics for very large absorbing matrices can be easily computed without going through tedious theoretical calculations to find explicit mathematical expressions.

Event-recall in probability learning

Subjects run in a probability learning (PL) task for 8 days with 200 trials a day were also asked to recall on each trial, n , the event that occurred on some previous trial, n-x . Values of x were varied to emphasize early or late trials and ranged from 1 to 9. On the PL task all groups showed some overshooting of the event probability, π, but the greater effect was observed when S s were forced to attend to remote trials ( x > 7 on more than half the trials). Recall performance, P c , varied with π and x as expected but the crucial factor controlling P c was found to be the number of runs of events occurring between trials n and n-x . There was substantial evidence that when S s do not remember the event on trial n-x , they guess with probability π. An attempt to develop a threshold model of recall with a bias parameter was not entirely successful.

An all-or-none model for noncorrection routines with elimination of incorrect responses.

Abstract A model is developed for paired-associate learning in which the correct response is not identified on error trials (the noncorrection procedure). The principal assumptions made are (1) that subjects can learn to eliminate error responses and (2) that response elimination and associative processes are all-or-none. A number of statistics including the mean learning curve, sequential statistics, the distributions of the trial of the last error, the total number of errors, and the number of error trials between the kth and (k + 1)st success are presented for the three-response case.

Matrix Analysis of Identifiability of Some Finite Markov Models.

Methods developed by Bernbach [1966] and Millward [1969] permit increased generality in analyses of identifiability. Matrix equations are presented that solve part of the identifiability problem for a class of Markov models. Results of several earlier analyses are shown to involve special cases of the equations developed here. And it is shown that a general four-state chain has the same parameter space as an all-or-none model if and only if its representation with an observable absorbing state is lumpable into a Markov chain with three states.

Event memory in probability learning

On each of 1600 trials of a probability learning experiment, Ss not only predicted the next event but also recalled the event on the xth past trial (x=l,...,9). The frequency distribution of x, f(x), was the main independent variable. Ss did not match event probability but seemed to overshoot as a function of f(x).