Predicting favorable and unfavorable consequences of perceptual learning: worsening and the peak shift

Abstract

Discrimination learning can cause improved and worsened ability to perceive differences. This subsequently affects how stimuli are associated with meanings and behaviors. Here, human listeners were trained with frequency-modulated (FM) tonal sweeps (500-1000 Hz) in a paradigm where one FM rate (8.29 octaves per second) required a 'Target' response, while a rate either slower (5.76 octaves per second) or faster (11.94 octaves per second) required a 'Non-Target' response. Training led to a shift in 'Target' responding along the FM rate dimension away from the 'Target' in a direction opposite the trained 'Non-Target'. This peak shift was paralleled by an asymmetry in acuity along the FM rate dimension in an untrained ABX task (a.k.a. match-to-sample). Performance improved relative to pre-training on trials where the 'Target' was contrasted with stimuli nearer the trained 'Non-Target'. Performance worsened on trials containing stimuli displaced along the FM dimension further from the trained 'Non-Target'. A connectionist model of perceptual learning containing non-associative representational modification and associative-based task-specific reweighting was able to simulate behavior. Simulations generated novel testable predictions regarding peak shift and worsening as a result of discrimination learning. Data have theoretical and practical consequences for predicting trends in the generalization of learned behaviors and modifiable perceptual acuities.

Modelling generalisation gradients as augmented Gaussian functions

“…In addition to the ease of interpretation of its parameters, the Gaussian function is appropriate for modelling generalisation for other reasons. Gaussian functions often play the role of the activation function that is assumed, partly for neurophysiological plausibility, in associative learning and neural network models that have been used to simulate generalisation gradients (e.g., Blough, 1975; McLaren & Mackintosh, 2002; Wisniewski, 2017). In other words, they have a conceptual place in associative models that assume graded activation of elements representing the stimulus dimension.…”

Learning to detect auditory signals in noise: Active top-down selection and stable change in signal representations.

“…In recent years, it has become increasingly acknowledged that even when listening performance between two conditions is matched (cf., the trained vs. novel frequencies observed here), the effort needed to achieve that level of performance can be drastically different (for review, see Francis & Love, 2020). Though most of this work has been conducted with speech in comparisons between clear and degraded speech signals, we have shown in several studies that the same can be seen in psychoacoustic tasks with nonspeech (Wisniewski, 2017a(Wisniewski, , 2017bWisniewski et al, 2018). Taking the preceding mechanic example, this would be apparent if while listening for a familiar engine sound, the mechanic would have more cognitive spare capacity left to do another task than if she was listening for an unfamiliar sound.…”

“…There are also differences among models proposing long-term plasticity in signal representations. The difference that has received the most debate is whether perceptual learning occurs in the read-out connections from static sensory representations to decision/response areas, or in sensory representations themselves (for extensive review, see Hall, 1991;McGann, 2015;Wisniewski, 2017aWisniewski, , 2017b. That we see changes in the AEP when no decision is being made could be interpreted as evidence for changes in sensory representations.…”

Benefits of fading in perceptual learning are driven by more than dimensional attention

“…Similar methods have been used successfully in memory research (for review, see [81]), and could make it possible to design empirically validated training regimens, without an exhaustive corpus of behavioral studies. This work could also be informative in establishing training procedures that limit “worsening” in generalization (i.e., when training hurts performance with novel stimuli; [22,45,82–84]), or that optimally benefits learning for specific types of perceptual input.…”