Normative modeling of auditory memory for natural sounds

Auditory memory is critical for nearly every daily task we perform involving sounds. However, memory is also imperfect, and the reasons for memory errors remain poorly understood. We probed auditory memory using a sound recognition experiment: sounds were presented serially, with some repeating, and participants judged whether each stimulus had occurred previously. Recognition was well above chance but decreased with the number of intervening stimuli between the first and second presentations of repeated sounds (the interstimulus interval; ISI). Some sounds were mistakenly judged to be novel when they had in fact repeated, and others were mistakenly judged to have repeated when they did not. To understand these effects in normative terms, we replicated them with auditory texture stimuli, whose internal representation is believed to be well explained by statistics of standard auditory filter cascades modeled after the auditory system. We then developed an ideal observer model using this statistical representation. The model encodes experiences as memory traces whose noise grows with time, and then decides whether a subsequent stimulus is more likely to have come from the distribution of stimuli implied by these memory traces or from the prior distribution over all textures. We estimated the prior from statistics measured from a large set of natural textures. The model qualitatively mirrored human performance trends and replicated human judgements on an item-by-item basis, predicting both false memories (stimuli incorrectly judged to have repeated) and misses (repeated stimuli that were incorrectly judged to be novel). The results suggest that memory errors can be predicted to a large extent from the statistics of natural stimuli and provide a framework for understanding the successes and failures of memory for natural stimuli.