Abstract

P. Tseng, D. J. Berg, M. Yoshida, I. G. M. Cameron, S. E. Boehnke, R. A. Marino, T. Ikeda, R. Kato, K. Takaura, L. Itti, T. Isa, D. P. Munoz, Deployment of visual attention and of the eyes differentiate observer populations, In: Annual Meeting of the Human Frontier Science Program (HFSP), Berlin, Germany, Jul 2008.

Abstract: In our complex visual world, it is critical to pay attention to the right places to detect life-threatening events, to accomplish tasks, and to explore new environments. Our attention is attracted by salient objects and events (bottom-up) and is controlled by the task at hand (top-down). A computational model of bottom-up attentional selection mechanisms (Itti & Koch, 2001) was developed based on the functionality of primary visual cortex (V1), which computes low-level features (color contrast, intensity contrast, orientation, flicker, and motion) from visual scenes, and detects locations where these features significantly differ from their surround. The model computes a saliency map which predicts hot spots that draw peopleÕs attention. To investigate the relative contributions of bottom-up and top-down in humans and monkeys, we tracked gaze of different groups of observers, in different but related projects supported by our grant: 15 control human children, 12 control human elderly, 6 Attention Deficit Hyperactivity Disorder (ADHD) patients, 4 Fetal Alcohol Spectrum Disorder (FASD) patients, 9 Parkinson's Disease (PD) patients, 4 normal rhesus monkeys, 3 macaque monkeys with unilateral lesion of V1, and 3 normal macaque monkeys) while they freely viewed videos. Using the saliency model, we quantified how bottom-up attentional selection mechanisms may differ between populations, based on the correlation between their eye-movement traces and the saliency maps, and using support-vector-machine classifiers to compare this correlation between populations. Leave-one-out was used to train and test the classifiers. The first experiment differentiated PD patients from control elderly with 95.24% correctness. The second differentiated ADHD, FASD from control children with 97% correctness. Further investigations comparing humans (young adults) and monkeys showed that the salience of locations which all monkeys looked at simultaneously was significantly higher than for humans (t-test, p<10-10). Moreover, the saliency model was used to quantify specific eye movement deficits in the monkeys with unilateral V1 lesion, demonstrating significant (t-test, p<10-5) albeit surprisingly small differences. Our results demonstrate that the balance between bottom-up and top-down mechanisms is greatly influenced by diseases or dysfunction, and can be quantified by the saliency model. This can serve not only to further our basic understanding of vision and attention, but also as a screening or diagnosis tool for clinical applications.

Themes: Model of Top-Down Attentional Modulation, Computational Modeling, Human Psychophysics, Model of Bottom-Up Saliency-Based Visual Attention, Human Eye-Tracking Research, Medical Research