Computational and Experimental Evaluation of the Attentional Blink: Testing the Simultaneous Type Serial Token Model

Computational and Experimental Evaluation of the Attentional Blink: Testing The Simultaneous Type Serial Token Model Bradley Wyble (B.Wyble@kent.ac.uk) and Howard Bowman (H.Bowman@kent.ac.uk) Computing Laboratory, University of Kent, Canterbury, Kent, CT2 7NF UK Abstract The Full ST 2 model: Basic Principles A reduced version of the Simultaneous Type Serial Token Model is presented. This model embraces two theories of temporal attention: Chun and Potter’s two-stage theory and Kanwisher’s types-tokens theory. We explain the proposed model and show how it reproduces key data from the Attentional Blink phenomenon. In addition, we verify experimentally predictions arising from the model. Introduction Rapid Serial Visual Presentation (RSVP) has been used extensively to explore how humans deploy attention over time (Chun & Potter, 1995; Kanwisher, 1987; Raymond, Shapiro, & Arnell, 1992; Weighelsgartner & Sperling 1987) Prominent amongst RSVP tasks is the Attentional Blink (AB), in which a decline in performance on a 2 nd target (hereafter the T2) is observed when it is presented within between 100 and 500ms of the offset of a 1 st target (hereafter the T1) (Chun & Potter, 1995; Raymond et al., 1992). Until recently, theoretical debates centred on informal explanations of the blink, which included the 2- stage model (Chun & Potter, 1995) and the Interference theory (Shapiro et al., 1997). These informal explanations have proved valuable in focusing formulation of experimental questions. However, the maturity of the field now makes it ripe for computational modelling. Accordingly, a number of models have recently been proposed (review: Bowman & Wyble, submitted). The current paper focuses on the STST (Simultaneous Type Serial Token) model (Bowman & Wyble, submitted; Bowman, Wyble, & Barnard, 2004), for which we also use the shorthand ST 2 . This paper reports the development of a reduced version of the approach (called the reduced ST 2 model), which abstracts from some of the implementation details of the full model. Importantly though, the new model remains consistent with the theoretical principles that underlie the earlier ST 2 incarnation. The focus of both ST 2 models is the letters-in-digits paradigm (Chun & Potter, 1995), in which the subject’s task is to report the two letter targets placed in a stream of digit distractors. This experiment can be viewed as a canonical AB methodology since no task switch is involved between T1 and T2. The task originally used in (Raymond et al., 1992) involved a task switch, and is therefore more complex. This paper serves to describe the reduced ST 2 model, its predictions, and, finally experimental verification of these predictions. Before introducing the new model, we briefly review necessary elements of the full ST 2 model. The full ST 2 model began as a rate-coded neural network elaboration of the theoretical two-stage model (Chun, 1997; Chun & Potter, 1995). The central idea behind their conception of RSVP processing is that the 1 st stage can represent multiple items in parallel, but only for a short time (several hundred milliseconds). The 2 nd stage is required for consolidation into a memory store that can persist until the end of the trial. However the 2 nd stage is limited in its ability to process multiple items concurrently. Therefore, in order to limit interference in the 2 nd stage, a gate is shut that denies entry to subsequent items. While waiting to be processed, these items are vulnerable to decay if they have been masked. We agree with Chun (1997) that this model can be well implemented using the types and tokens framework described by Kanwisher (1987). In her theory, types represent possible kinds of items, devoid of context. In contrast, tokens represent memories that a given item was encountered, i.e. episodic (instance specific) information about the occurrence of an item. A token can be bound to any combination of types within its domain. In the case of the letters-in-digits task, types would include all of the letters and digits, and one or more tokens would be assigned to represent the occurrence of the targets within an RSVP stream. The other key difference between types and tokens is that the latter are strictly sequential in nature, in that only one token may be in the process of binding at any one time. Hence, the name of our approach: Simultaneous Type, Serial Token. Types can be considered analogous to stage 1 of the two stage theoretical model, while our token implementation is the analogue of stage 2. Stage 1. The full ST 2 model represented each potential item (i.e. each type) in an RSVP stream with a unique node that would be re-activated if the item occurred twice. We implemented a series of layers that represented steps of visual processing. At the 1 st layer, distractors backward masked targets via inhibitory connections. Remaining activation from these masked traces reached the task- selective layer, at which the task demand system emphasized targets. Stage 2. Target(s) in stage-1 could activate a token gate in stage-2, while strong lateral inhibition ensured only one token could be active at any time. The remainder of the token layer implemented dynamics, which insured, that after being active for a sufficient amount of time (approximately 200-300 msec), a token gate would be