Naslov
A Neurodynamical Model of How Prior Knowledge Influences Visual Perception

Autori
Domijan, Dražen ; Marić, Mateja

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
10th Context in Psychology Symposium - Book of Abstracts / Domijan, Dražen ; Marić, Mateja - Ilok, 2017, 6-6

Skup
10th Context in Psychology

Mjesto i datum
Ilok, Hrvatska, 16.06.2017. - 18.06.2017

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
visual perception, adaptive resonance theory, neural networks

Sažetak
Many behavioral studies showed that prior knowledge can directly influence visual perception (Firestone & Scholl, 2016). For example, Goldstone (1995) found that categorization of simple objects into letters and numerals influences perception of their color. During training, letters were arbitrarily associated with red hues and numerals were associated with violet hues. In the test phase, letter L and numeral 8 were presented in the same hue that was midway between red and violet. Results showed that participants judged L to be redder and numeral 8 to be more violet than it was. In the current work, we offer an explanation of the observed finding based on the well-established neurocomputational theory known as the adaptive resonance theory (ART). The ART neural network was designed to solve the problem of catastrophic forgetting during learning in non- stationary environment. In the ART, stability of learning is achieved by matching bottom-up sensory signals with top-down expectations. Resonant state that corresponds with conscious perception develops in the network when the bottom-up and top-down signals are closely aligned. On the other hand, mismatch produces global reset signal that clears the traces of erroneous top-down expectations. Therefore, prior knowledge can influence conscious perception only when it already closely matches with sensory signals. To explain the findings of Goldstone (1995), we employed two ART modules, one encoding object shape and another encoding object color. These modules communicate via inter-ART associative map enabling both modules to influence each other by reading out expectations consistent with their current activity. To set up connection weights in the network, we employed fuzzy ARTMAP algorithm (Carpenter et al., 1992). Next, we proposed that color ART module outputs to the decision or response selection stage. Further, we assumed that this output is noisy and we modelled it by the Gaussian function. Computer simulations showed that observed alternation in perceived color arises from the activity bias that develops in the response selection stage although the color ART module encodes veridical hue. We conclude that category knowledge does not influence perception directly but rather it bias response selection.

Izvorni jezik
Engleski

Znanstvena područja
Psihologija

POVEZANOST RADA