A Connexionist Model of the Attentional Blink Effect During a Rapid Serial Visual Presentation Task

We introduce a connectionist model that reproduces the attentional blink effect during a rapid serial visual presentation task. The model is composed of two layers, a competitive one that acts as an identification layer and a recurrent one that acts like short term memory where the main mechanisms is the presence of an inhibition gate and a neural fatigue. Simulations show that the model generates data that are as variables as the one obtained from human participants and the mean performance is identical to the performance obtained by human participants. Introduction During visual scanning, the eyes gaze at a given area then make a saccade to another area. Between the saccades, they receive brief images. How the brain can process such a flow of information is studied in the laboratory using the Rapid Serial Visual Presentation (RSVP) task. In this experimental setting, stimuli are presented in rapid succession, usually, 6 to 20 stimuli per second, at a same spatial location. Within the stream, there is usually a target which is marked by a different attribute (e.g. color) and the task is to identify it. Thus, the RSVP can be seen as a visual search without saccade. The attentional blink (AB) effect is characterized by a decrease in performance for a second target when a first target has been identified (T2|T1). The performance for recalling a second target will decrease if it is presented within 200 to 500 milliseconds of the first. However, if the second target is presented next to the first target, there is no decrease in performance. This phenomenon is called lag of 1 sparing effect. Figure 1 shows a typical AB curve with human participants obtained from our lab. Many independent models have been postulated to explain the AB phenomenon.The first explanation is given by Raymond, Shapiro & Arnell (1992) and is based on the postulate of an inhibition process. When a first target is perceived, the perceptual system is inhibited to avoid confusion with subsequent items. However, the inhibition is slow to start, so that if a second target is next to the first, it will be processed along the way. Figure 1: Results with human participants in the AB task from our lab. Shapiro, Ward & Duncan (1997) challenged this explanation by showing that the second target could facilitate the processing of a third target. If perception was inhibited, there should be no facilitation, according to Raymond & al. (1992). Another explanation was proposed by Chun & Potter (1995) who postulated a two-stage model. The first stage operates rapidly and decides which of the stimuli is sent to the second stage for encoding. The second stage is slow and can process only one item at a time. Meanwhile, the second target is left in a waiting stage. Thus, as time passes the probability of encoding the second target correctly decreases. However, if the second target is presented next to the first target, it will be encoded Finally, the most recent explanation comes from Jolicoeur (1998). He proposed a modified version of the two-stage model by adding a third stage. According to Jolicoeur, the first and last stages are the perception and recall stage respectively. The second stage is used to encode the targets in short term memory and is characterize by its limited capacity (1 item at a time). This 50 60 70 80 90 1 2 3 4 5 6 7 8 Lag % o f co rr ec t re sp o n se s