What has been learned from computational models of attention

There are few fields where neurocomputational modelling 1 is as necessary as in that of attention — a multifaceted and 2 elusive process that is at the very core of cognition and con3 sciousness (James, 1890). While descriptive theories of at4 tentional functions, relying on a plethora of metaphors (spot5 light/zoomlens, increased-gain, biased-competition), abound, 6 computational models that are simple enough to promote un7 derstanding of the phenomena and make testable predictions are 8 relatively scarce. As models are often criticised for their ability 9 to ‘fit everything’, imposing some degree of constraint is essen10 tial. Neurocomputational models can answer this challenge by 11 taking on neurophysiological constraints and by addressing not 12 only behavioural but also physiological data. I start with some 13 functional considerations of the attention process, followed by 14 examples of models that help to explicate these processes and 15 some suggestions for future work. 16 The most central characteristic of attention, common to 17 all its subtypes, is a limited capacity bottleneck (Broadbent, 18 1958; James, 1890). The nature of this bottleneck, however, 19 is likely to vary with the various types. One aspect of the 20 bottleneck (the late one) involves the selection of information 21 for transfer (consolidation) to (in) short-term memory (Chun 22 & Potter, 1995; Duncan, 2006). Target selection is thought to 23 involve a type of top-down control characterised by biased24 competition (Desimone & Duncan, 1995) and is subject to 25 capacity limitations. The bottleneck is demonstrated in multiple 26 target paradigms (Duncan, 1980), where a strong interference 27 takes place (unlike in single target with multiple distractors 28 paradigms). In the attentional blink (AB) paradigm (Raymond, 29 Shapiro, & Arnell, 1992), for example, the detection of the 30 second target (T2) is considerably depressed when it follows a 31 previous (detected) target (T1) up to intervals of about 500 ms. 32 Interestingly, when the interval between the two targets is only 33