Connectionist Modeling of Normal and Impaired Cognitive Functions

summary of the ongoing research in Prof. Carlo Umilta`s research group

(Psychology Dept., Univ. of Padova).

Connectionist Modeling of Normal and Impaired Cognitive Functions

University of Padova, Italy

Department of General Psychology

www.psico.unipd.it

Cognitive modeling has become a fundamental step for understanding brain functions and the processes underlying human behavior and performance. In the recent years, the introduction of formal modeling techniques (i.e., neural networks), contrasting the loosely formulated box-and-arrows models, has driven a revolution in the theoretical thinking, by forcing scientists to focus attention to difficult (and yet basic) problems such as the nature of the cognitive processes and of the representations on which these processes operate. Connectionist models have proved to be more fruitful than traditional approaches in many psychological areas, providing an account of learning and emergence of "rule-governed" behavior from network of simple units and connections. Our main interest is in developing neural network models of human cognitive functions, in particular in the fields of language and attention; the aim is to investigate how these models can account from data regarding the skilled performance of normal subjects and the impaired performance of individuals that suffered from brain injury (see, e.g., Zorzi et al., in press). Besides this main topic, we have been using neural networks in the field of medical diagnosis (Paccagnella et al., 1996).

For the purpose of modeling cognitive functions, we use a variety of network topologies and learning algorithms. The choice of network topology is driven by a careful analysis of the domain and of the input-output representations that are relevant for the task. In particular, the latter aspect determines whether a recoding process is necessary for learning the mapping: if this is the case, an intermediate layer is needed to provide internal representations of the input patterns. This can be either a hidden layer (e.g., Zorzi et al., in press; Zorzi, 1994a) or a Kohonen layer (e.g., Erba et al., in press). In other cases, however, the network can be simply two-layered (i.e., with input and output layers only), and the task is learnt simply through one-shot hebbian learning (e.g., Zorzi & Umilta`, 1995).

Input and output representations may vary widely for different tasks. However, they generally consist in distributed patterns of activation over a set of units that encode "microfeatures", that is a collection of basic, low-level features that describes a given input stimulus. For example, the visual representation of objects may consist of a "structural description", that is a set of features describing the object's components (e.g., in terms of geometrical shapes) and the spatial relations between components (see, e.g., Erba et al., in press).

The performance of a network model is usually compared to human skilled performance both in terms of mean reaction times for producing a given response and in terms of error percentages. A stochastic component (i.e., noise) in the model produces response variability that often allows a better match to human performance. In some cases, however, the model must

also be able to account for the impaired performance of patients that suffered from brain damage; therefore, the model is "lesioned" after training (e.g., by removing weights or units) and its residual abilities are compared to that of the patients. Finally, when the model provides a good match to the existing data, it is very important to derive from the model's structure and/or behavior some new predictions that can be later tested experimentally.

Relevant publications:

Zorzi, M., Houghton, G., & Butterworth, B. (in press). Two routes or one in

reading aloud? A connectionist "dual-process" model. Journal of

Experimental Psychology: Human Perception and Performance.

Erba, A., Zorzi, M., & Umilta`, C. (in press).

Categorizzazione percettivab e semantica in soggetti umani e reti neurali. Giornale Italiano di Psicologia.

Paccagnella, F., Zorzi, M., & Sartori, G. (1996).

Predicting head injuried patient's outcome: A neural expert system. In: M. Witten (Ed.),

Computational Medicine, Public Health and Biotechnology: Building a Man in

the Machine (Part 3). Singapore: World Scientific Publishing.

Umilta`, C., & Zorzi, M. (1996).

Learning and attention in S-R compatibility. In B. Hommel & W. Prinz (Eds.), Theoretical issues in

stimulus-response compatibility. Amsterdam: North Holland.

Zorzi, M., & Umilta`, C. (1995).

A computational model of the Simon effect.

Psychological Research, 58, 193-205.

Zorzi, M. (1994a).

Reti neurali modulari e simulazione di elaborazioni

cognitive complesse. Giornale Italiano di Psicologia, 21( 2), 205-220.

Zorzi, M. (1994b).

Connessionismo e percezione: Un'analisi critica.

Giornale Italiano di Psicologia, 21 (3), 329-336.

Download this Application Summary in Word97 format.
If you don't have Word97....Download the Word Viewer.