@inproceedings{Navalpakkam_Itti04jsnc,
  author = {V. Navalpakkam and L. Itti},
  title = {A mathematical framework for the design and analysis of feature biasing strategies},
  abstract = {Aim: Given a target and a set of distractors, we wish to find a desirable feature biasing strategy that will render the target most salient and suppress inteference from the distractors. That is, we wish to find how bottom-up features such as color, intensity, orientation can be biased in a top-down manner so as to enable quick detection of the target amidst distractors. Desirable feature biasing strategy: If a feature is present in the target and absent in the distractor, promoting it can further boost the target's salience. Since it already leads to a pop-out that is optimal performance, biasing is not required. If a feature is present in both the target and the distractor in an equal amount, then biasing the feature cannot yield a performance gain. To investigate this, we designed a SAME type distractor (see figure 2). Whereas, if the feature is present in a less amount in the target as compared to the distractor, then suppressing this feature can boost the target s salience relative to the distractor. To investigate this, we designed a MORE type distractor that contained the target s feature in a greater amount, and another extreme case - a NEW type distractor that contained a new feature absent in the target. Design and analysis of experiment: To test whether humans use the desirable strategy, we designed search arrays containing the target and all 3 types of distractors and measured the relative number of fixations on each type of distractor (see figure 1 for the list of target and distractors and their difference, and figure 3 for sample search arrays). We tested 7 subjects on 600 trials over a period of 5 days each. Paired t tests over the combined data of all subjects and also for individual subjects supported our hypothesis that humans fixate more on SAME type distractors than MORE (p value = 7.6471e-11) and NEW type distractors (p value = 1.4125e-09), indicating that they used the desirable strategy of suppressing all features as each feature was present in a lesser amount in the target than the distractors. Conclusion: We have provided a mathematical framework for the analysis and design of experiments to test various feature biasing strategies, to determine a desirable strategy and to test whether humans use that strategy.},
  booktitle = {Proc. 11th Joint Symposium on Neural Computation (JSNC'04), Los Angeles, California},
  month = {May},
  year = {2004},
  type = {td;psy;eye},
  file = {http://iLab.usc.edu/publications/doc/Navalpakkam_Itti04jsnc.pdf},
  review = {abs/conf}
}