Time as the supervisor: Unsupervised learning of classification of natural auditory stimuli via Slow Feature Analysis*

A fundamental goal of sensory systems is to extract sensory information from the environment and convert it into perceptual representations. As the brain cannot simply transduce and represent all of the information present in the environment, sensory systems must select features of the stimuli to encode.

One way a sensory system could perform this feature selection is by encoding particular statistical regularities in the environment. One statistical regularity of natural auditory stimuli is that they tend to have low temporal modulation; i.e. the powers of the frequencies that comprises natural stimuli tend to change slowly over time.  It is unknown whether such slow temporal regularities are sufficient to enable learning and perception of auditory object classes.

To test this idea, we adapted an unsupervised temporal learning algorithm, Slow Feature Analysis (SFA), to extract the auditory features that change most slowly over time.  We then used this algorithm to evaluate the hypothesis that extracting these slowly varying features will capture both intra- and inter-class stimulus variance of rhesus macaque vocalizations.  We found that (1) pairs of vocalizations in the SFA-generated feature space were linearly separable; (2) this feature space is robust to clutter (noise) in the training data set; and (3) this feature space captures enough variability for the classification of novel exemplars. Together, our results suggest that if the brain can extract the slow temporal features from auditory stimuli, it may be sufficient for and underlie important components of perception.