Human and Computer Learning: An Experimental Study

ALEXANDRE TSALLIS; CONSTANTINO TSALLIS; AGLAÉ MAGALHAES; FRANCISCO TAMARIT

Complexus

Karger

Lugar: Paris; Año: 2004 vol. 1 p. 181 - 189

1424-8492

Simple memorizing tasks have been chosen such as a binary code on a 5 × 5 matrix. After the establishment of an appropriate protocol, the codified matrices were individually presented to 150 university students (conveniently preselected) who had to memorize them. Multiple presentations were offered seeking perfect performance verified through the correct reproduction of the code. We measured the individual percentage error as a function of the number of successive presentations, and then averaged over the examined population. The learning curve thus obtained decreases (almost monotonically) until becoming virtually zero when the number of presentations attains six. A computer simulation for a similar task is available which uses a two-level perceptron on which an algorithm was implemented allowing for some degree of globality or nonlocality (technically referred to as entropic nonextensivity within a current generalization of the usual, Boltzmann-Gibbs, statistical mechanics). The degree of nonextensivity is characterized by an index q, such that q = 1 recovers the usual, extensive, statistical mechanics, whereas q ≠ 1 implies some degree of nonextensivity. In other words, q ? 1 is a (very sensitive) measure of globality (gestalt perception or learning). The computer curves fit well the human result for q . 1.02. It has been verified that even extremely small departures of q from unity lead to strong differences in the learning curve. Our main observation is that, for the very specific learning task on which we focus here, humans perform similarly to slightly nonextensive perceptrons. In addition to this experiment, some preliminary studies were done concerning the human learning of ambiguous images (based on figure-background perception). In spite of the complexity of drawing conclusions from such a comparison, some generic trends can be established. Moreover, the enormous and well-known difficulty for computationally defining semantic, hierarchic and strategic structures reveals clear-cut differences between human and machine learning.