Computational Modeling of Human Attention

Although “everyone knows what attention is” (James, 1890) , the nature of attention remains one of the central puzzles in science. Recent findings in psychology and brain imaging have increasingly suggested that it is better to view attention not as a unitary faculty of the mind but as a complex organ system subserved by multiple interacting neuronal networks in the brain (Posner & Raichle, 1994) . At least three such attentional networks, for alerting (achieving and maintaining an alert state in preparation for coming stimuli), orienting (selectively focusing on one or a few items out of many candidate ones), and executive control (monitoring and resolving conflicts in planning, decision-making, error detection, and overcoming habitual actions), have been identified. Considerable functional neuroimaging evidence has shown activities of these areas highly correlate with the essential functions of attention. A experimental paradigm, named Attentional Networks Test (ANT), was developed by Dr. Jin Fan and colleagues (e.g., Fan et al, 2002, JCNS) in an attempt to systematically explore and evaluate the cognitive functions of the networks as well as the underlying brain foundations (via EEG/fMRI). 

A sketch of the three attentional networks. The alerting network consists of the frontal and parietal cortical regions particularly of the right hemisphere. The orienting network consists of parts of the superior and inferior parietal lobe, frontal eye fields and such subcortical areas as the superior colliculus of the midbrain and the pulvinar and reticular nucleus of the thalamus. The executive control network includes the midline frontal areas (especially the anterior cingulate cortex, or ACC), lateral prefrontal cortex, and the basal ganglia.

How the distinct attentional networks interact and work together to achieve the functions of attention is not clear. Computational modeling has the potential to provide important computational links that lead from neuroimaging evidence to behavioral data. Seeking such principled computational links is also a challenge that is facing the field of cognitive neuroscience in general. In this project, we attempt to explore such a link by developing computational models of human attentional networks, at both neural networks level and cognitive/symbolic level.

The neural networks level model (please contact me at Hongbin.Wang@uth.tmc.edu for a copy of the model) is developed in the framework of PDP++, adopting the algorithm of leabra (local, error-driven and associative, biologically realistic algorithm) (O'Reilly & Munakata, 2000). The structure of the model is shown below. This model contains modules for all the three attentional networks. In addition, it contains modules for perception (visual input and primary visual cortex), object recognition (object pathway), and response (output). The networks are connected in such a way that they conform to the known functional an anatomical constraints described above as much as possible. The model works as follows. When a cue comes on, the primary visual cortex module is activated, which in turn triggers the alerting network. This cue-induced alerting affects later stimulus processing because the alerting network will remain excited for a while which will activate the orienting network in general causing it to become ready for the incoming stimulus. In addition, when the cue is a spatial one (i.e., a cue that indicates where the target stimulus is to appear), it will further make the corresponding sub-region of the orienting network even more excited. This occurs because the orienting network adopts a retinotopy-based spatial representation of the environment. This extra excitation in the sub-region of the orienting network will facilitate the corresponding stimulus processing in the object pathway network, due to the connections between them. This accounts for the orienting effect. Finally, note that it is the object pathway network that is responsible for the arrow direction detection. When the incongruent stimulus (e.g., a left arrow flanked by four right arrows) is presented, the object pathway network may propose different responses, which compete for the final expression in the output network. The executive control network then activates making the center arrow defeating the flankers. This is where the executive control attention plays a role.

   

A sketch of the computational model of attentional networks. The modules are connected bi-directionally (except the one from visual input to primary visual cortex).

The cognitive/symbolic level model is developed in the framework of Act-R (Anderson & Lebiere, 1998), a production-based cognitive architecture. With the add-on perceptual-motor module, Act-R is capable of interacting with the same ANT software that real human subjects interact with. You can request a copy of the model by sending me an email.

The table below summarizes the preliminary modeling results, along with empirical behavioral results adopted from Fan et al 2002 (note that the double cue condition is not listed).  It is clear that both models fit the behavioral data reasonably well, with the correlation coefficient >= 0.94.

The multilevel computational models developed in this project not only provides a fine-grained computational explanation of how the attentional networks work and lead to cognition but also permits us to make quantitative predictions about how attention will behavior in various normal and abnormal conditions. Perhaps most importantly, by developing detailed computational models of the same phenomenon and the same task at multiple levels, we are now capable of systematically comparing and contrasting them. We are one step closer to a better understanding of how neural activities in the brain are related to cognitive behavior.