Reverse hierarchies and sensory learning

Abstract

Revealing the relationships between perceptual representations in the brain and mechanisms of adult perceptual learning is of great importance, potentially leading to significantly improved training techniques both for improving skills in the general population and for ameliorating deficits in special populations. In this review, we summarize the essentials of reverse hierarchy theory for perceptual learning in the visual and auditory modalities and describe the theory's implications for designing improved training procedures, for a variety of goals and populations.

Figure 1

Schematic of the local-to-global processing hierarchies. (a) An example of the visual hierarchy (adapted from Hochstein & Ahissar 2002) and (b) an example of the auditory hierarchy (see another example in Shamma 2008).

Figure 2

Performance of speech perception in noise when subjects need to do semantic processing. (a) The experimental paradigm. Subjects listened to the words, masked by speech noise, through headphones. Shortly after the onset of the oral presentation, a word was visually flashed on the screen for 500 ms. Subjects had to decide whether the word they heard (in this case, ‘day’) semantically matches the visually presented word (‘dream’), and press the correct button (in this case, ‘match’). (b–c) In order to measure usefulness of low-level information, words and noise were presented in two different conditions (used in separate blocks). In one, the same word and noise were presented to both ears simultaneously. In the other, the word was presented to the two ears with inverse phase. The latter condition maximizes the information that the auditory system can derive from the interaural difference, which is calculated at low levels. The difference in threshold between these two conditions is the benefit from the low-level binaural information (denoted as filled bars in the figure). The maximal binaural benefit that can be obtained by the system if it uses all low-level information fully can be calculated. This number (the ‘ideal listener’ benefit, see Nahum et al. 2008) is denoted by open bars in the figure. (b) Local information is only partially used (measured benefit<ideal listener) when discriminating between phonologically similar words. (c) Local information is fully used (measured benefit=ideal listener benefit) when discriminating between words with no phonological overlap. Adapted from Nahum et al. 2008.

Figure 3

Speech repetition in noise. Under these conditions, the benefit from using low-level binaural information is similar and optimal both when the words are perceptually (phonologically) (a) similar and (b) different. The absolute difficulty of this task was similar to that of the semantic task depicted in figure 2, as noted by the fact that the diotic thresholds, which contain no specific low-level information, were the same (data not shown). Adapted from Nahum et al. 2008.

Figure 4

A crude scheme of the expected learning dynamics under two different protocols. In one (black line) subjects are trained in blocks, each containing a narrow range of stimuli. In the other (dashed grey line), the two stimulus ranges are mixed within the same block (‘interleaved’ blocks 1 and 2 contain the same mixed stimuli). According to RHT, when each stimulus range is learned separately, performance on each range becomes gradually better than when all cues are learned together. The blocked condition enables access to the most informative low-level population for each cue, whereas the interleaved condition keeps the performer at higher, broader levels. Thus, asymptotic performance is worse using interleaved stimulus ranges. Yet, it is reached faster and yields greater generalization to untrained parameters. Thus, as shown for block 3, when a third range of stimuli is introduced, blocked training requires re-learning, whereas performance following interleaved training is largely generalized (i.e. does not have to be re-learned).

Figure 5

Illustration of the crowding effect. Fixate on the dot (filled circles) on the left of (a,b) and try to identify the central letter on the right. (a) Isolated letter—identification is easy when only one letter is presented. (b) Crowding—identification is much harder when other letters surround the middle letter.

Figure 6

A schematic of the top-down cascade of learning across hierarchical layers. For simplicity, three levels of a hierarchical structure are featured. (a) Initial learning phase. Since learning follows reverse hierarchy order, the first connection strengthened is the presumed informative connection which feeds into the higher node (red thick line; connecting between nodes c and B). (b) Subsequent learning phase. With more training on specific conditions, learning proceeds backwards to the most informative lower level population, strengthening the connection between lower node 4 and middle node c (path denoted in red). As a result, the response in upper level node B is strengthened, but since node c also feeds into upper nodes A and C, their connections are also somewhat strengthened (dotted red lines). Therefore, following this phase, learning is partially transferred to higher level contexts that use the trained lower level features.