Gift from statistical learning: Visual statistical learning enhances memory for sequence elements and impairs memory for items that disrupt regularities

Statistical Learning in Vision

Vision and learning have long been considered to be two areas of research linked only distantly. However, recent developments in vision research have changed the conceptual definition of vision from a signal-evaluating process to a goal-oriented interpreting process, and this shift binds learning, together with the resulting internal representations, intimately to vision. In this review, we consider various types of learning (perceptual, statistical, and rule/abstract) associated with vision in the past decades and argue that they represent differently specialized versions of the fundamental learning process, which must be captured in its entirety when applied to complex visual processes. We show why the generalized version of statistical learning can provide the appropriate setup for such a unified treatment of learning in vision, what computational framework best accommodates this kind of statistical learning, and what plausible neural scheme could feasibly implement this framework. Finally, we list the challenges that the field of statistical learning faces in fulfilling the promise of being the right vehicle for advancing our understanding of vision in its entirety. Expected final online publication date for the Annual Review of Vision Science, Volume 8 is September 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

Visual object recognition is facilitated by temporal community structure

Humans and others primates are highly attuned to temporal consistencies and regularities in their sensory environment and learn to predict such statistical structure. Moreover, in several instances, the presence of temporal structure has been found to facilitate procedural learning and to improve task performance. Here we extend these findings to visual object recognition and to presentation sequences in which mutually predictive objects form distinct clusters or "communities." Our results show that temporal community structure accelerates recognition learning and affects the order in which objects are learned ("onset of familiarity").

Semantically predictable input streams impede gaze-orientation to surprising locations

When available, people use prior knowledge to predict dimensions of future events such as their location and semantic features. However, few studies have examined how multi-dimensional predictions are implemented, and mechanistic accounts are absent. Using eye tracking, we evaluated whether predictions of target-location and target-category interact during the earliest stages of orientation. We presented stochastic series so that across four conditions, participants could predict either the location of the next target-image, its semantic category, both dimensions, or neither. Participants observed images in absence of any task involving their semantic content. We modeled saccade latencies using ELATER, a rise-to-threshold model that accounts for accumulation rate (AR), variance of AR over trials, and variance of decision baseline. The main findings were: 1) AR scaled with the degree of surprise associated with a target's location; 2) predictability of semantic-category hindered saccade latencies, suggesting a bottleneck in implementing joint predictions; 3) saccades to targets that satisfied semantic expectations were associated with greater AR-variance than saccades to semantically-surprising images, consistent with a richer repertoire of early evaluative processes for semantically-expected images. Predictability of target-category also impacted gaze pre-positioning prior to target presentation. The results indicate a strong interaction between foreknowledge of object location and semantics during stimulus-guided saccades, and suggest statistical regularities in an input stream can also impact anticipatory, non-stimulus-guided processes.

Effects of tDCS on visual statistical learning

Visual statistical learning describes the encoding of structure in sensory input, and it has important consequences for cognition and behaviour. Higher-order brain regions in the prefrontal and posterior parietal cortices have been associated with statistical learning behaviours. Yet causal evidence of a cortical contribution remains limited. In a recent study, the modulation of cortical activity by transcranial direct current stimulation (tDCS) disrupted statistical learning in a spatial contextual cueing phenomenon; supporting a cortical role. Here, we examined whether the same tDCS protocol would influence statistical learning assessed by the Visual Statistical Learning phenomenon (i.e., Fiser and Aslin, 2001), which uses identity-based regularities while controlling for spatial location. In Experiment 1, we employed the popular exposure-test design to tap the learning of structure after passive viewing. Using a large sample (N = 150), we found no effect of the tDCS protocol when compared to a sham control nor to an active control region. In Experiment 2 (N = 80), we developed an online task that was sensitive to the timecourse of learning. Under these task conditions, we did observe a stimulation effect on learning, consistent with the previous work. The way tDCS affected learning appeared to be task-specific; expediting statistical learning in this case. Together with the existing evidence, these findings support the hypothesis that cortical areas are involved in the visual statistical learning process, and suggest the mechanisms of cortical involvement may be task-dependent and dynamic across time.

Neural Mechanisms of Memory Enhancement and Impairment Induced by Visual Statistical Learning

Prior research has reported that the medial temporal, parietal, and frontal brain regions are associated with visual statistical learning (VSL). However, the neural mechanisms involved in both memory enhancement and impairment induced by VSL remain unknown. In this study, we examined this issue using event-related fMRI. fMRI data from the familiarization scan showed a difference in the activation level of the superior frontal gyrus (SFG) between structured triplets, where three objects appeared in the same order, and pseudorandom triplets. More importantly, the precentral gyrus and paracentral lobule responded more strongly to Old Turkic letters inserted into the structured triplets than to those inserted into the random triplets, at the end of the familiarization scan. Furthermore, fMRI data from the recognition memory test scan, where participants were asked to decide whether the objects or letters shown were old (presented during familiarization scan) or new, indicated that the middle frontal gyrus and SFG responded more strongly to objects from the structured triplets than to those from the random triplets, which overlapped with the brain regions associated with VSL. In contrast, the response of the lingual gyrus, superior temporal gyrus, and cuneus was weaker to letters inserted into the structured triplets than to those inserted into the random triplets, which did not overlap with the brain regions associated with observing the letters during the familiarization scan. These findings suggest that different brain regions are involved in memory enhancement and impairment induced by VSL.

A common probabilistic framework for perceptual and statistical learning

System-level learning of sensory information is traditionally divided into two domains: perceptual learning that focuses on acquiring knowledge suitable for fine discrimination between similar sensory inputs, and statistical learning that explores the mechanisms that develop complex representations of unfamiliar sensory experiences. The two domains have been typically treated in complete separation both in terms of the underlying computational mechanisms and the brain areas and processes implementing those computations. However, a number of recent findings in both domains call in question this strict separation. We interpret classical and more recent results in the general framework of probabilistic computation, provide a unifying view of how various aspects of the two domains are interlinked, and suggest how the probabilistic approach can also alleviate the problem of dealing with widely different types of neural correlates of learning. Finally, we outline several directions along which our proposed approach fosters new types of experiments that can promote investigations of natural learning in humans and other species.

From the structure of experience to concepts of structure: How the concept "cause" is attributed to objects and events

The pervasive presence of relational information in concepts, and its indirect presence in sensory input, raises the question of how it is extracted from experience. We operationalized experience as a stream of events in which reliable predictive relationships exist among random ones, and in which learners are naïve as to what they will learn (i.e., a statistical learning paradigm). First, we asked whether predictive event pairs would spontaneously be seen as causing each other, given no instructions to evaluate causality. We found that predictive information indeed informed later causal judgments but did not lead to a spontaneous sense of causality. Thus, event contingencies are relevant to causal inference, but such interpretations may not occur fully bottom-up. A second question was how such experience might be used to learn about novel objects. Because events occurred either around or involving a continually present object, we were able to distinguish objects from events. We found that objects can be attributed causal properties by virtue of a higher-order structure, in which the object's identity is linked not to the increased likelihood of its effect, but rather, to the predictive structure among events, given its presence. This is an important demonstration that objects' causal properties can be highly abstract: They need not refer to an occurrence of a sensory event per se, or its link to an object, but rather to whether or not a predictive relationship holds among events in its presence. These learning mechanisms may be important for acquiring abstract knowledge from experience. (PsycINFO Database Record (c) 2019 APA, all rights reserved).

Predictable Events Enhance Word Learning in Toddlers

Sensitivity to the predictability of the environment supports young children's learning in many domains [1, 2], including language [3-6]; perception [7, 8]; and the processing of objects, space, and time [1, 9]. Predictable regularities allow observers to generate expectations about upcoming events and to learn from violations of those expectations [10, 11]. Given the benefits of detecting both predictable and unpredictable events, a key question concerns which types of input facilitate learning in young children. In the current research, we assessed the effects of predictability on toddlers' word learning by embedding word-learning moments within events that were either predicted or violated predictions. 2-year-olds observed a continuous visual sequence in which novel objects were revealed from one of four locations in a predictable spatiotemporal pattern (1, 2, 3, 4). Objects were then labeled either during events that were predicted by the sequence (1, 2, 3, 4) or events that violated the sequence (1, 2, 3, 2). Results from two studies revealed better word learning for objects labeled during predictable events than objects labeled during unpredictable events. These findings suggest that predictable events create advantageous learning moments for toddlers, with implications for the role played by predictable input in early development.

Selection History Modulates Working Memory Capacity

Recent studies have shown that past selection history affects the allocation of attention on target selection. However, it is unclear whether context-driven selection history can modulate the efficacy of attention allocation on working memory (WM) representations. This study tests the influences of selection history on WM capacity. A display of one item (low load) or three/four items (high load) was shown for the participants to hold in WM in a delayed response task. Participants then judged whether a probe item was in the memory display or not. Selection history was defined as the number of items attended across trials in the task context within a block, manipulated by the stimulus set-size in the contexts with fewer possible stimuli (4-item or 5-item context) or more possible stimuli (8-item or 9-item context) from which the memorized content was selected. The capacity measure (i.e., the K measure) was estimated to reflect the number of items that can be held in WM. Across four behavioral experiments, the results revealed that the capacity was significantly reduced in the context with more possible stimuli relative to the context with fewer possible stimuli. Moreover, the reduction in capacity was significant for high WM load and not observed when the focus was on only a single item. Together, these findings indicate that context-driven selection history and focused attention influence WM capacity.