We examine a model of the process of stimulus identification, which assumes that complex visual or auditory stimuli are represented as vectors in a multidimensional perceptual space, and which postulates a simple probabilistic decision process based on the geometric structure of the perceptual space. We present evidence from several conditions of an identification task that human observers engage in a continuing, dynamic process in which dimension salience weights are tuned to optimize identification performance. In addition, we verify the reliability of the INDSCAL multidimensional scaling procedure in deriving the geometric structure of the observers' perceptual space for the set of visual spectrograms used in our identification tasks. We also present evidence supporting an assumption of dimensional decomposability made in the decision process. Finally, we observe that the model is successful in accounting for approximately 90 percent of the variance in individual confusion matrices, averaged over 18 observers x conditions. (Author)
