Naslov
Visual feature integration with amplitude maximization

Autori
Domijan, Dražen

Izvornik
Review of Psychology (1330-6812) 8 (2001), 1-2; 31-40

Vrsta, podvrsta i kategorija rada
Radovi u časopisima, članak, znanstveni

Ključne riječi
Feature integration; Object recognition; Neural network; Lateral connections; Presynaptic inhibition

Sažetak
A neural model of visual feature integration is proposed based on presynaptic inhibition of excitatory feedback connections. The same activity level or amplitude of corresponding nodes represents features that belong to the same object. This is achieved by spreading of activation from strongly activated nodes to weakly activated nodes but not in reverse. Spreading is controlled by presynaptic inhibition, which prevents unbounded activity growth. The model’ s representational capacity is far greater than in models based on temporal synchrony, it is equally applicable to the static and moving stimuli and it can represent hierarchical groupings. A network may operate as a short-term storage, which allows simultaneous feature integration over space and time. Top-down signals from higher visual centers can influence the operation of the network either through direct excitatory input or indirectly through inhibitory interneurons. With direct input, model behavior is consistent with psychophysical data on object-based attentional selection and curve tracing task. Indirect influences are able to provide flexible task-dependent feature integration and prevent accidental bindings that are consequences of intrinsic properties of some scenes.

Izvorni jezik
Engleski

Znanstvena područja
Psihologija

POVEZANOST RADA