Statistical Learning in a Multisensory World

Abstract

In this mini-review, the general paradigm used for Statistical Learning (SL) was traced and the main theoretical debate of brain modular vs. centralized processing was introduced, mainly from the evidence of multisensory interaction that is documented in the literature. Moreover, the time course of SL has been concisely delineated. Finally, a survey of the neuronal exploration of SL was given, although the endeavor in revealing neural mechanism is insufficient so far.

Keywords: Statistical learning; Cross-modality; Development; Modular

Introduction

As active learners, we rely on a combination of experience-independent and experience-dependent mechanisms to extract information from the environment. Statistical Learning (SL) has been studied as a mechanism by which people automatically discover patterns in the environment through experience. SL starts remarkably early in the progress of human life-span development. For instance, 8-month-old infants are capable of extracting serial-order information after only 2 min of listening experience [1].

It is the brain's capacity to detect statistical regularities in the environment, by operating complex perceptual and cognitive manipulations to obtain object recognition [2,3], event identification [4-7], and even language acquisition [8,9]. The information of the sensory events is usually presented sequentially, given by a specific sensory modality or a combination of modalities (such as auditory and visual modalities). The efficiency of SL differs among different sensory modalities. In general, auditory modality displays a quantitative learning advantage compared with vision and touch [4,5]. The disparities in sensory processing have been commonly recognized as sensory dominance ever since 1980 [10]. To substantiate the sensory dominance/difference in SL, a paradigm of artificial grammar has been developed and applied extensively in a large body of experimental explorations. A typical procedure of experiment goes as follows: observers are required to make 'match' or 'mismatch' discrimination of the two presented stimuli sequences (both are visual, auditory or tactile sequences), in which the presentation orders of spatial locations for a visual square, or the pitches of an auditory sequence containing multiple beeps were aligned by the given predefined grammar. Using the artificial grammar protocol, a number of studies have demonstrated that the auditory modality appears to have an advantage in the processing of sequential input, including low-level temporal processing tasks and pattern or rhythm discrimination, while touch modality is adept at processing both sequential and spatial input, though it is not at the same level of proficiency as either audition or vision [4-6,11].

Supra-modal or modular processing?

A central debate in SL concerns whether learners encode the regularities with an abstract or stimulus-specific representation. Modular theories hypothesize that SL is accomplished by mechanisms particular to the domain in question [12-14]. However, it has been widely observed various domains and many sorts of species that could attune to probabilistic patterns in the environment [15-18] suggesting a supra-modal representation of SL from phylogenetic perspective. To resolve this debate, people usually investigate whether there is a transfer of learning (benefits) from one stimulus set to another, independent of perceptual features of the stimuli or the sensory modality. For example, whether there is a transfer from the dimension (visual shapes) to another dimension (auditory pitch), namely cross-modally. Alternatively, one may examine how the cross-modal relationships influence simultaneous learning of multimodal input streams. For the latter, Seitz et al. [19] found statistical learning is a modality-independence process in which observers could extract concurrent, multiple (auditory vs. visual) statistical patterns equally, supporting the view of modular (modality-independence) processing [19]. In contrast, Mitchel and Weiss [6] presented both auditory and visual streams simultaneously or asynchronously, with variable predictability (transitional probabilities) between audio and visual elements and asked the observers to segment the boundaries of the audio and visual triplets (Figure 1). The results suggest that learners were able to extract multiple statistical regularities across modalities provided that there is some degree of cross-modal coherence [6], favoring a supra-modal abstract representation.

Figure 1:  Stimuli pattern adapted from Mitchel and Weiss [6]. Two sequences of auditory music tones and artificial visual figures were presented. The correspondence of each tone and figure is manipulated to the predefined artificial grammar.