Change Detection and Visual Classification: Two Sides of the Same Coin

Change is certain, but the what, the where and the when are unpredictable. E.g. approaching animals need to be detected promptly, and also classified as friend or foe. What is the best way to carry out detection and discrimination jointly? How do humans do it? We compared subjects' performance in three tasks: (1) detection with unpredictable stimulus onset time, (2) discrimination with known stimulus onset time, and (3) dual task: discrimination with unpredictable stimulus onset time. Four subjects participated in a novel random-dot motion discrimination task, where a trial begins with incoherent random-dot motion and coherent motion appears after a random delay. For different tasks, observers were required to detect the onset of the signal and/or discriminate the motion direction. In our stimuli motion directions are not necessarily opposite, so we could vary independently the difficulty of discrimination and detection by controlling the motion strength (% coherence) and the motion directions. Our data show that discrimination is more difficult when signal onset is uncertain, suggesting that the additional detection task imposes an extra cost on the system. We developed three models of joint detection and discrimination: an optimal model (CD: classify-then-detect), and two simplified versions of the optimal model (DCP: detection-classification-in-parallel, and DCS: detection-classification-in-series). We compare the predictions of the three models to human data, and find that CD does not fit the data well, as it has no flexibility to switch off the discrimination component, and will overestimate the difficulty of the detection-only task. In addition, DCS produces the most consistent parameter estimates among different tasks. DCS is sub-optimal because it discards signal for discrimination until detection. Thus our findings indicate that the visual system may not be optimal from a purely information-processing point of view. Meeting abstract presented at VSS 2016