Naslov
Modeling the Influence of Visual Perception on Conceptual Processing

Autori
Šetić, Mia ; Domijan, Dražen

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

Izvornik
Eleventh International Conference on Cognitive and Neural Systems : Proceedings / - Boston (MA) : Boston University, 2007, 23-23

Skup
International Conference on Cognitive and Neural Systems (11 ; 2007)

Mjesto i datum
Boston (MA), Sjedinjene Američke Države, 16.05.2007. - 19.05.2007

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Concepts; Neural Model; Knowledge Representation; Language Understanding; Perception

Sažetak
Traditional theories of knowledge representation assume that perception is not important for understanding abstract thoughts and language. On the other hand, Barsalou (1999) developed a theory of perceptual symbol systems which claims that understanding the meaning of concepts necessarily involves reactivation of perceptual experiences associated with them. Reactivation of memorized perceptual experience is called perceptual simulation. Predictions of the theory were tested in numerous behavioral and neuroimaging experiments (Gibbs, 2006). For instance, Zwaan et al. (2002) showed that listening to a sentence influences picture recognition performance. When the word mentioned in a sentence matches with the object presented in a picture, recognition is faster compared to the situation where there is no such match. Šetić and Domijan (in press) showed that meaning of words is processed faster when words were presented in a location where we usually observe denoting objects. Barsalou’ s theory is defined verbally and it is not clear what computational and neural mechanisms may subserve perceptual simulation. We suggest that adaptive resonance theory (ART) is appropriate framework for understanding interaction between perception and conception (Carpenter & Grossberg, 2003). In ART, stability of conceptual learning is achieved by matching output of perceptual processing with top-down signals or expectations from learned concepts. When bottom-up and top-down codes are sufficiently similar, resonance occurs, which stores pattern of activation in a long-term memory. On the other hand, when there is a mismatch between perceptual and conceptual codes, reset signal temporary disables currently active concept node and search for a new concept node begins. In order to explain results from behavioral experiments we extended the ART architecture by introducing motor or output neurons which receive signals from concept nodes. Motor neurons accumulate evidence in real-time and fire when they reach threshold for activating motor response. Difference in reaction times in the experiment of Zwaan et al. (2002) is accounted for by the fact that listening to sentence activates concept nodes for words mentioned in a sentence. Concept node prepares or primes visual perceptual nodes by subthreshold activation. If subsequent picture presentation matches with conceptual expectation, reaction time is facilitated. If subsequent picture does not match with expectation, reset is issued until appropriate concept node is found. Reset and search among concept nodes prolong reaction time as observed experimentally. Difference in reaction times in the experiment of Šetić and Domijan (in press) is accounted for by facilitation due to the difference in strength of bottom-up synaptic connections between locations in perceptual map and concept nodes. For instance, top locations can have stronger connections with concepts like birds while bottom locations can have stronger connections with concepts like dog and cat. Results of these computer simulations show that ART could be viewed as a neural basis for perceptual simulation.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Informacijske i komunikacijske znanosti, Psihologija

POVEZANOST RADA