Sequence Encoders Enable Large-Scale Lexical Modeling: Reply to Bowers and Davis (2009)

Sibley, Kello, Plaut, and Elman (2008) proposed the sequence encoder as a model that learns fixed-width distributed representations of variable-length sequences. In doing so, the sequence encoder overcomes problems that have restricted models of word reading and recognition to processing only monosyllabic words. Bowers and Davis (2009) recently claimed that the sequence encoder does not actually overcome the relevant problems, and hence it is not a useful component of large-scale word-reading models. In this reply, it is noted that the sequence encoder has facilitated the creation of large-scale word-reading models. The reasons for this success are explained and stand as counterarguments to claims made by Bowers and Davis.     

Implicit sequence learning in a search task

This study investigated the effects of selection demands on implicit sequence learning. Participants in a search condition looked for a target among seven distractors and responded on the target identity. The responses followed a deterministic sequence, and sequence learning was compared to that found in two control conditions in which the targets were presented alone, either at a central location or over a series of unpredictable locations. Sequence learning was obtained in all conditions, and it was equivalent for the two variable location conditions, regardless of the perceptual demands. Larger effects of learning were observed in the central location, both on the indirect measures and on the measures taken from a cued-generation task. The expression of learning decreased selectively in this condition when the sequence validity was reduced over a test block. These results are consistent with the claims that implicit and explicit learning are mixed in this central condition and that implicit learning is not affected by selection difficulty.     

What is learned about fragments in artificial grammar learning? A transitional probabilities approach

Learning local regularities in sequentially structured materials is typically assumed to be based on encoding of the frequencies of these regularities. We explore the view that transitional probabilities between elements of chunks, rather than frequencies of chunks, may be the primary factor in artificial grammar learning (AGL). The transitional probability model (TPM) that we propose is argued to provide an adaptive and parsimonious strategy for encoding local regularities in order to induce sequential structure from an input set of exemplars of the grammar. In a variant of the AGL procedure, in which participants estimated the frequencies of bigrams occurring in a set of exemplars they had been exposed to previously, participants were shown to be more sensitive to local transitional probability information than to mere pattern frequencies.     

Learning what from where: Effects of spatial regularity on nonspatial sequence learning and updating

The current study examined the influence of redundant stimulus features on our ability to build and update representations of our environment. We hypothesized that our ability to process redundant spatial features would speed our ability to adapt to changing nonspatial regularities. Using a computerized version of the children's game “rock–paper–scissors”, undergraduates were instructed to win as often as possible against a computer opponent. The computer's plays were repeating sequences of five choices that were presented either with spatial regularity (i.e., “rock” would always appear on the left, “paper” in the middle, and “scissors” on the right) or without spatial regularity (i.e., the items were equally likely to appear in any of the three locations). Once participants learned a sequence, the computer switched to a different sequence without participants being informed that a switch had occurred. Redundant spatial regularity improved a participant's ability both to learn sequences of plays and to update their plays to reflect new computer sequences. Our results suggest that our perceptual system is sensitive to redundant spatial stimulus features and that this information can improve learning and updating.     

Using Dynamic Field Theory to extend the embodiment stance toward higher cognition

The embodiment stance emphasizes that cognitive processes unfold continuously in time, are constantly linked to the sensory and motor surfaces, and adapt through learning and development. Dynamic Field Theory (DFT) is a neurally based set of concepts that has turned out to be useful for understanding how cognition emerges in an embodied and situated system. We explore how the embodiment stance may be extended beyond those forms of cognition that are closest to sensorimotor processes. The core elements of DFT are dynamic neural fields (DNFs), patterns of activation defined over different kinds of spaces. These may include retinal space and visual feature spaces, spaces spanned by movement parameters such as movement direction and amplitude, or abstract spaces like the ordinal axis along which sequences unfold. Instances of representation that stand for perceptual objects, motor plans, or action intentions are peaks of activation in the DNFs. We show how such peaks may arise from input and are stabilized by intra-field interaction. Given a neural mechanism for instantiation, the neuronal couplings between DNFs implement cognitive operations. We illustrate how these mechanisms can be used to enable architectures of dynamic neural fields to perform cognitive functions such as acquiring and updating scene representations, using grounded spatial language, and generating sequences of actions. Implementing these DFT models in autonomous robots demonstrates how these cognitive functions can be enacted in embodied, situated systems.     

One-year retention of general and sequence-specific skills in a probabilistic,serial reaction time task

Procedural skills such as riding a bicycle and playing a musical instrument play a central role in daily life. Such skills are learned gradually and are retained throughout life. The present study investigated 1-year retention of procedural skill in a version of the widely used serial reaction time task (SRTT) in young and older motor-skill experts and older controls in two experiments. The young experts were college-age piano and action video-game players, and the older experts were piano players. Previous studies have reported sequence-specific skill retention in the SRTT as long as 2 weeks but not at 1 year. Results indicated that both young and older experts and older non-experts revealed sequence-specific skill retention after 1 year with some evidence that general motor skill was retained as well. These findings are consistent with theoretical accounts of procedural skill learning such as the procedural reinstatement theory as well as with previous studies of retention of other motor skills.     

儿童青少年协作促进的发展

旨在考察不同年龄的个体, 在不同回忆序列和学习材料上协作促进的发展情况。采用3(回忆序列:协作—个人—个人〔CII〕、个人—协作—个人〔ICI〕、个人—个人—个人〔III〕)× 5(年龄:9、11、14、17、20)× 2(材料:图片、文字)的被试间设计。结果发现:(1)各年龄个体在CII和ICI两种回忆序列上均出现了协作促进, 且两种序列条件下被试的最终个人回忆成绩无差异;(2)另外, 17岁个体的协作促进量显著高于9岁、11岁、14岁和20岁, 其他各年龄之间无差异;(3)图片和文字的协作促进量无差异。这说明小组协作回忆的顺序对协作促进没有太大影响, 17岁个体协作促进的效果最好。     

Stimulus dependence and cross-modal interference in sequence learning

A central issue in sequence learning is whether learning operates on stimulus-independent abstract elements, or whether surface features are integrated, resulting in stimulus-dependent learning. Using the serial reaction-time (SRT) task, we test whether a previously presented sequence is transferrable from one domain to another. Contrary to previous artificial grammar learning studies, there is mapping between pre- and posttransfer stimuli, but contrary to previous SRT studies mapping is not obvious. In the pre-transfer training phase, participants face a dot-counting task in which the location of the dots follows a predefined sequence. In the test phase, participants face an auditory SRT task in which the spatial organization of the response locations is either the same as spatial sequence in the training phase, or not. Sequence learning is compared to two control conditions: one with a non-sequential random-dot counting in the training phase, and one with no training phase. Results show that sequential training proactively interferes with later sequence learning, regardless of whether the sequence is the same or different in the two phases. Results argue for the existence of a general sequence processor with limited capacity, and that sequence structures and sequenced elements are integrated into a single sequential representation.     

Motor sequence learning performance in Parkinson's disease patients depends on the stage of disease

It is still unclear, whether patients with Parkinson's disease (PD) are impaired in the incidental learning of different motor sequences in short succession, although such a deficit might greatly impact their daily life. The aim of this study was thus to clarify the relation between disease parameters of PD and incidental motor learning of two different sequences in short succession. Results revealed that the PD patients were able to acquire two sequences in short succession but needed more time than healthy subjects. However, both the severity of axial manifestations, as assessed on a subsection of the Unified Parkinson's Disease Rating Scale III (UPDRS III) and the Hoehn and Yahr score, and the levodopa-equivalent dose (LED) were negatively correlated with the sequence learning performance. These findings indicate that, although PD patients are able to learn two sequences in short succession, they need more time and their overall sequence learning performance is strongly correlated with the stage of disease.     

信息生命科学——对于信息量生命、本体信息演化规律特征的探索

信息生命科学不是生物学计算机化,不是生命科学,也不是生命信息科学,而是信息生命科学。