Representing word context effect in Chinese character perception — an application of information theory

The perception of a letter in the context of a word is easier than in the context of a random letter sequence. It appears that our knowledge about words can influence our perception process. McClelland and Rumelhart (1981) propose an interactive activation model to account for the interaction between our knowledge about words and our visual input. They use their model to explain how these interactions facilitate perception. In their account, word context effect is a constant independent of the identity of the words. In this paper, we propose the use of informatin theory to quantify word context effect. In this way, the strength of word context effect will depend on the identity of the words. We apply the method to quantify word context effect in Chinese words. This knowledge is encoded in an artificial neural network using the interactive activation and competition model. The network is used to recognize Chinese characters and we are able to achieve a high recognition rate.     

Two models of the recognition and detection of texture-defined letters compared

We quantified texture segregation by measuring psychophysically the percentage correct detection scores for each of a set of 10 texture-defined (TD) letters using the temporal two-alternative forced choice method, and at the same time quantified spatial discrimination of the TD form of measuring psychophysically the percentage correct letter recognition scores for the 10 letters. Ten levels of task difficulty were created by adding noise dots to the texture patterns. The resulting psychophysical data were used to test and compare models of the detection and recognition of texture-defined letters. Each model comprised a sequence of physiologically plausible stages in early visual processing. Each had the same first, second and third stages, namely linear orientation-tuned spatial filters followed by rectification and smoothing. Model 1 had only one non-linear stage. Model 2 had two non-linear stages. In model 2 the second non-linear stage was cross-orientation inhibition. This second non-linear stage enhanced the texture borders by, in effect, comparing textures at different locations in the texture pattern. In both models, the last stage modelled either letter detection or letter recognition. Letter recognition was modelled as follows. We passed a given letter stimulus through the first several stages of a model and, in 10 separate calculations, cross-correlated the output with a template of each of the 10 letters. From these 10 correlations we obtained a predicted percentage correct letter recognition score for the given letter stimulus. The predicted recognition scores closely agreed with the experimental data at all 10 levels of task difficulty for model 2, but not for model 1. We conclude that a borderenhancing algorithm is necessary to model letter recognition. The letter-detection algorithm modelled detection of part of a letter (a single letter stroke) in terms of the signal-to-noise ratio of a letter-segment detector. The predicted letter detection scores fitted the data closely for both models.     

Recognition of partly occluded patterns: a neural network model

 Human beings are often able to read a letter or word partly occluded by contaminating ink stains. However, if the stains are completely erased and the occluded areas of the letter are changed to white, we usually have difficulty in reading the letter. In this article I propose a hypothesis explaining why a pattern is easier to recognize when it is occluded by visible objects than by invisible opaque objects. A neural network model is constructed based on this hypothesis. The visual system extracts various visual features from the input pattern and then attempts to recognize it. If the occluding objects are not visible, the visual system will have difficulty in distinguishing which features are relevant to the original pattern and which are newly generated by the occlusion. If the occluding objects are visible, however, the visual system can easily discriminate between relevant and irrelevant features and recognize the occluded pattern correctly. The proposed model is an extended version of the neocognitron model. The activity of the feature-extracting cells whose receptive fields cover the occluding objects is suppressed in an early stage of the hierarchical network. Since the irrelevant features generated by the occlusion are thus eliminated, the model can recognize occluded patterns correctly, provided the occlusion is not so large as to prevent recognition even by human beings. Received: 21 February 2000 / Accepted in revised form: 11 September 2000     

A New Font,Specifically Designed for Peripheral Vision,Improves Peripheral Letter and Word Recognition,but Not Eye-Mediated Reading Performance

Reading speed is dramatically reduced when readers cannot use their central vision. This is because low visual acuity and crowding negatively impact letter recognition in the periphery. In this study, we designed a new font (referred to as the Eido font) in order to reduce inter-letter similarity and consequently to increase peripheral letter recognition performance. We tested this font by running five experiments that compared the Eido font with the standard Courier font. Letter spacing and x-height were identical for the two monospaced fonts. Six normally-sighted subjects used exclusively their peripheral vision to run two aloud reading tasks (with eye movements), a letter recognition task (without eye movements), a word recognition task (without eye movements) and a lexical decision task. Results show that reading speed was not significantly different between the Eido and the Courier font when subjects had to read single sentences with a round simulated gaze-contingent central scotoma (10° diameter). In contrast, Eido significantly decreased perceptual errors in peripheral crowded letter recognition (-30% errors on average for letters briefly presented at 6° eccentricity) and in peripheral word recognition (-32% errors on average for words briefly presented at 6° eccentricity).     

The Lazy Visual Word Form Area: Computational Insights into Location-Sensitivity

In a recent study, Rauschecker et al. convincingly demonstrate that visual words evoke neural activation signals in the Visual Word Form Area that can be classified based on where they were presented in the visual fields. This result goes against the prevailing consensus, and begs an explanation. We show that one of the simplest possible models for word recognition, a multilayer feedforward network, will exhibit precisely the same behavior when trained to recognize words at different locations. The model suggests that the VWFA initially starts with information about location, which is not being suppressed during reading acquisition more than is needed to meet the requirements of location-invariant word recognition. Some new interpretations of Rauschecker et al.''s results are proposed, and three specific predictions are derived to be tested in further studies.     

Bayesian action-perception computational model: interaction of production and recognition of cursive letters

In this paper, we study the collaboration of perception and action representations involved in cursive letter recognition and production. We propose a mathematical formulation for the whole perception-action loop, based on probabilistic modeling and bayesian inference, which we call the Bayesian Action-Perception (BAP) model. Being a model of both perception and action processes, the purpose of this model is to study the interaction of these processes. More precisely, the model includes a feedback loop from motor production, which implements an internal simulation of movement. Motor knowledge can therefore be involved during perception tasks. In this paper, we formally define the BAP model and show how it solves the following six varied cognitive tasks using bayesian inference: i) letter recognition (purely sensory), ii) writer recognition, iii) letter production (with different effectors), iv) copying of trajectories, v) copying of letters, and vi) letter recognition (with internal simulation of movements). We present computer simulations of each of these cognitive tasks, and discuss experimental predictions and theoretical developments.     

Early-visual factors in letter confusions

For the purpose of quantifying models of letter recognition, similarities are often specified in terms of stimulus properties. In this paper, an approach is advocated based on similarities between internal letter representations or internal letter images, i.e. it is argued that optical and retinal factors play a more prominent role in letter confusions than is usually assumed. To illustrate this, letter images were calculated on the basis of earlier experimentally determined point spread functions (Blommaert et al., Spatial Vision 2, 99-115, 1987). Next, data on confusion matrices from Bouma (Vision Res. 11, 459-474, 1971) were taken to evaluate different measures which might be useful for quantifying similarities between internal letter representations. In the analysis of experimental data, Luce's (In: Handbook of Mathematical Psychology, 1963) choice model was used. It was found that if similarities were expressed in terms of differences between image contours, a fair first order approximation of Bouma's experimental results could be formulated (overall correlation coefficient of 0.95). Other measures like correlations between spatial frequency spectra of letter images were found to be less successful. The method used provides a means to relate quantitatively stimulus features and optical and early-visual factors to letter confusions.     

A Four-Stage Hybrid Model for Hydrological Time Series Forecasting

Hydrological time series forecasting remains a difficult task due to its complicated nonlinear, non-stationary and multi-scale characteristics. To solve this difficulty and improve the prediction accuracy, a novel four-stage hybrid model is proposed for hydrological time series forecasting based on the principle of ‘denoising, decomposition and ensemble’. The proposed model has four stages, i.e., denoising, decomposition, components prediction and ensemble. In the denoising stage, the empirical mode decomposition (EMD) method is utilized to reduce the noises in the hydrological time series. Then, an improved method of EMD, the ensemble empirical mode decomposition (EEMD), is applied to decompose the denoised series into a number of intrinsic mode function (IMF) components and one residual component. Next, the radial basis function neural network (RBFNN) is adopted to predict the trend of all of the components obtained in the decomposition stage. In the final ensemble prediction stage, the forecasting results of all of the IMF and residual components obtained in the third stage are combined to generate the final prediction results, using a linear neural network (LNN) model. For illustration and verification, six hydrological cases with different characteristics are used to test the effectiveness of the proposed model. The proposed hybrid model performs better than conventional single models, the hybrid models without denoising or decomposition and the hybrid models based on other methods, such as the wavelet analysis (WA)-based hybrid models. In addition, the denoising and decomposition strategies decrease the complexity of the series and reduce the difficulties of the forecasting. With its effective denoising and accurate decomposition ability, high prediction precision and wide applicability, the new model is very promising for complex time series forecasting. This new forecast model is an extension of nonlinear prediction models.     

字母位置对左右视野的识别影响

目的:考查左右视野中不同位置字母的辨别力。方法:采用探测刺激的方法对左右视野中字母位置的识别差异进行了研究。结果:通过SPSS15.0进行重复测量方差分析得出:①不同视野的错误率主效应显著(F=5.98,P<0.05);②探测位置的主效应非常显著(F=15.39,P<0.01);③不同视野与探测字母位置的交互作用非常显著(F=11.60,P<0.001)。结论:①位置错误多于字母错误,并且位置错误中所报告的字母更靠近中央凹。②左视野中的首字母有较差的辨别力,而右视野中的尾字母有较差的辨别力。③左视野中存在知觉系统的限制,而不是词汇识别策略中大脑左右半球的差异。④对词汇识别中大脑左半球平行加工、右半球系列加工的说法提出了质疑。     

Letter Position Coding Across Modalities: The Case of Braille Readers

Background

The question of how the brain encodes letter position in written words has attracted increasing attention in recent years. A number of models have recently been proposed to accommodate the fact that transposed-letter stimuli like jugde or caniso are perceptually very close to their base words.

Methodology

Here we examined how letter position coding is attained in the tactile modality via Braille reading. The idea is that Braille word recognition may provide more serial processing than the visual modality, and this may produce differences in the input coding schemes employed to encode letters in written words. To that end, we conducted a lexical decision experiment with adult Braille readers in which the pseudowords were created by transposing/replacing two letters.

Principal Findings

We found a word-frequency effect for words. In addition, unlike parallel experiments in the visual modality, we failed to find any clear signs of transposed-letter confusability effects. This dissociation highlights the differences between modalities.

Conclusions

The present data argue against models of letter position coding that assume that transposed-letter effects (in the visual modality) occur at a relatively late, abstract locus.     

WEDeepT3: predicting type III secreted effectors based on word embedding and deep learning

Background: The type III secreted effectors (T3SEs) are one of the indispensable proteins in the growth and reproduction of Gram-negative bacteria. In particular, the pathogenesis of Gram-negative bacteria depends on the type III secreted effectors, and by injecting T3SEs into a host cell, the host cell’s immunity can be destroyed. The high diversity of T3SE sequences and the lack of defined secretion signals make it difficult to identify and predict. Moreover, the related study of the pathological system associated with T3SE remains a hot topic in bioinformatics. Some computational tools have been developed to meet the growing demand for the recognition of T3SEs and the studies of type III secretion systems (T3SS). Although these tools can help biological experiments in certain procedures, there is still room for improvement, even for the current best model, as the existing methods adopt hand-designed feature and traditional machine learning methods. Methods: In this study, we propose a powerful predictor based on deep learning methods, called WEDeepT3. Our work consists mainly of three key steps. First, we train word embedding vectors for protein sequences in a large-scale amino acid sequence database. Second, we combine the word vectors with traditional features extracted from protein sequences, like PSSM, to construct a more comprehensive feature representation. Finally, we construct a deep neural network model in the prediction of type III secreted effectors. Results: The feature representation of WEDeepT3 consists of both word embedding and position-specific features. Working together with convolutional neural networks, the new model achieves superior performance to the state-of-the-art methods, demonstrating the effectiveness of the new feature representation and the powerful learning ability of deep models. Conclusion: WEDeepT3 exploits both semantic information of k-mer fragments and evolutional information of protein sequences to accurately differentiate between T3SEs and non-T3SEs. WEDeepT3 is available at bcmi.sjtu.edu.cn/~yangyang/WEDeepT3.html.     

Incorporating Linguistic Knowledge for Learning Distributed Word Representations

Combined with neural language models, distributed word representations achieve significant advantages in computational linguistics and text mining. Most existing models estimate distributed word vectors from large-scale data in an unsupervised fashion, which, however, do not take rich linguistic knowledge into consideration. Linguistic knowledge can be represented as either link-based knowledge or preference-based knowledge, and we propose knowledge regularized word representation models (KRWR) to incorporate these prior knowledge for learning distributed word representations. Experiment results demonstrate that our estimated word representation achieves better performance in task of semantic relatedness ranking. This indicates that our methods can efficiently encode both prior knowledge from knowledge bases and statistical knowledge from large-scale text corpora into a unified word representation model, which will benefit many tasks in text mining.     

Erratum to: Functional Brain Organization of Global and Local Visual Perception: An Event-Related Potentials Study

Adult subjects were asked to recognize a hierarchical visual stimulus (a letter) while their attention was drawn to either the global or local level of the stimulus. Event-related potentials (ERP) and behavioral indices (reaction time and percentage of correct responses) were measured. An analysis of behavioral indices showed the global level precedence effect, i.e. the increase in a small letter recognition time when this letter is a part of incongruent stimulus. An analysis of ERP components showed level-related (global vs. local) differences in the timing and topography of the brain organization of perceptual processing and regulatory mechanisms of attention. Visual recognition at the local level was accompanied by (1) stronger activation of the visual associative areas (P _z and T ₆) at the stage of sensory features analysis (P1 ERP component), (2) involvement mainly of inferior temporal cortices of the right hemisphere (T ₆) at the stage of sensory categorization (P2 ERP component), and (3) involvement of prefrontal cortex of the right hemisphere at the stage of selection of the relevant features of the target (N2 ERP component). Visual recognition at the global level was accompanied by (1) pronounced involvement of mechanisms of early sensory selection (N1 ERP component), (2) prevailing activation of parietal cortex of the right hemisphere (P ₄) at the stage of sensory categorization (P2 ERP component) as well as at the stage of the target stimulus identification (P3 ERP component). We suggested that perception of the hierarchical stimulus at the global level is related primarily to the analysis of its spatial features in the dorsal visual system whereas the perception at the local level primarily involves an analysis of the object-related features in the ventral visual system.     

Temporal predictive codes for spoken words in auditory cortex

Humans can recognize spoken words with unmatched speed and accuracy. Hearing the initial portion of a word such as "formu…" is sufficient for the brain to identify "formula" from the thousands of other words that partially match. Two alternative computational accounts propose that partially matching words (1) inhibit each other until a single word is selected ("formula" inhibits "formal" by lexical competition) or (2) are used to predict upcoming speech sounds more accurately (segment prediction error is minimal after sequences like "formu…"). To distinguish these theories we taught participants novel words (e.g., "formubo") that sound like existing words ("formula") on two successive days. Computational simulations show that knowing "formubo" increases lexical competition when hearing "formu…", but reduces segment prediction error. Conversely, when the sounds in "formula" and "formubo" diverge, the reverse is observed. The time course of magnetoencephalographic brain responses in the superior temporal gyrus (STG) is uniquely consistent with a segment prediction account. We propose a predictive coding model of spoken word recognition in which STG neurons represent the difference between predicted and heard speech sounds. This prediction error signal explains the efficiency of human word recognition and simulates neural responses in auditory regions.     

Discrete emotion effects on lexical decision response times

Our knowledge about affective processes, especially concerning effects on cognitive demands like word processing, is increasing steadily. Several studies consistently document valence and arousal effects, and although there is some debate on possible interactions and different notions of valence, broad agreement on a two dimensional model of affective space has been achieved. Alternative models like the discrete emotion theory have received little interest in word recognition research so far. Using backward elimination and multiple regression analyses, we show that five discrete emotions (i.e., happiness, disgust, fear, anger and sadness) explain as much variance as two published dimensional models assuming continuous or categorical valence, with the variables happiness, disgust and fear significantly contributing to this account. Moreover, these effects even persist in an experiment with discrete emotion conditions when the stimuli are controlled for emotional valence and arousal levels. We interpret this result as evidence for discrete emotion effects in visual word recognition that cannot be explained by the two dimensional affective space account.     

Word learning and the cerebral hemispheres: from serial to parallel processing of written words

Reading familiar words differs from reading unfamiliar non-words in two ways. First, word reading is faster and more accurate than reading of unfamiliar non-words. Second, effects of letter length are reduced for words, particularly when they are presented in the right visual field in familiar formats. Two experiments are reported in which right-handed participants read aloud non-words presented briefly in their left and right visual fields before and after training on those items. The non-words were interleaved with familiar words in the naming tests. Before training, naming was slow and error prone, with marked effects of length in both visual fields. After training, fewer errors were made, naming was faster, and the effect of length was much reduced in the right visual field compared with the left. We propose that word learning creates orthographic word forms in the mid-fusiform gyrus of the left cerebral hemisphere. Those word forms allow words to access their phonological and semantic representations on a lexical basis. But orthographic word forms also interact with more posterior letter recognition systems in the middle/inferior occipital gyri, inducing more parallel processing of right visual field words than is possible for any left visual field stimulus, or for unfamiliar non-words presented in the right visual field.     

Comment on the Letter by A. Ben-Shaul: “Entropy,Energy, and Bending of DNA in Viral Capsids”

The conformational entropic penalty associated with packaging double-stranded DNA into viral capsids remains an issue of contention. So far, models based on a continuum approximation for DNA have either left the question unexamined, or they have assumed that the entropic penalty is negligible, following an early analysis by Riemer and Bloomfield. In contrast, molecular-dynamics (MD) simulations using bead-and-spring models consistently show a large penalty. A recent letter from Ben-Shaul attempts to reconcile the differences. While the letter makes some valid points, the issue of how to include conformational entropy in the continuum models remains unresolved. In this Comment, I show that the free energy decomposition from continuum models could be brought into line with the decomposition from the MD simulations with two adjustments. First, the entropy from Flory-Huggins theory should be replaced by the estimate of the entropic penalty given in Ben-Shaul’s letter, which corresponds closely to that from the MD simulations. Second, the DNA-DNA repulsions are well described by the empirical relationship given by the Cal Tech group, but the strength of these should be reduced by about half, using parameters based on the Rau-Parsegian experiments, rather than treating them as “fitting parameters (tuned) to fit the data from (single molecule pulling) experiments.”     

A Deep Learning Architecture for P300 Detection with Brain-Computer Interface Application

In this paper, a brain-computer interface (BCI) system for character recognition is proposed based on the P300 signal. A P300 speller is used to spell the word or character without any muscle movement. P300 detection is the first step to detect the character from the electroencephalogram (EEG) signal. The character is recognized from the detected P300 signal. In this paper, sparse autoencoder (SAE) and stacked sparse autoencoder (SSAE) based feature extraction methods are proposed for P300 detection. This work also proposes a fusion of deep-features with the temporal features for P300 detection. A SSAE technique extracts high-level information about input data. The combination of SSAE features with the temporal features provides abstract and temporal information about the signal. An ensemble of weighted artificial neural network (EWANN) is proposed for P300 detection to minimize the variation among different classifiers. To provide more importance to the good classifier for final classification, a higher weightage is assigned to the better performing classifier. These weights are calculated from the cross-validation test. The model is tested on two different publicly available datasets, and the proposed method provides better or comparable character recognition performance than the state-of-the-art methods.     

基于卷积神经网络和循环神经网络的环形RNA剪接位点识别研究

本文提出了一种基于卷积神经网络和循环神经网络的深度学习模型,通过分析基因组序列数据,识别人基因组中环形RNA剪接位点.首先,根据预处理后的核苷酸序列,设计了2种网络深度、8种卷积核大小和3种长短期记忆(long short term memory,LSTM)参数,共8组16个模型;其次,进一步针对池化层进行均值池化和最...     

Functional brain organization of global and local visual perception: an ERP study

Adult subjects were asked to recognize a hierarchical visual stimulus (a letter) while their attention was drawn to either the global or local level of the stimulus. Event-related potentials (ERP) and psychophysical indices (reaction time and percentage of correct responses) were measured. An analysis of psychophysical indices showed the global level precedence effect, i.e., the increase in a small letter recognition time when this letter is a part of incongruent stimulus. An analysis of ERP components showed level-related (global vs. local) differences in the timing and topography of the brain organization of perceptual processing and regulatory mechanisms of attention. Visual recognition at the local level was accompanied by (1) stronger activation of the visual associative areas (Pz and T6) at the stage of sensory features analysis (P1 ERP component), (2) involvement mainly of inferior temporal cortices of the right hemisphere (T6) at the stage of sensory categorization (P2 ERP component), and (3) involvement of prefrontal cortex of the right hemisphere at the stage of the selection of the relevant features of the target (N2 ERP component). Visual recognition at the global level was accompanied by (1) pronounced involvement of mechanisms of early sensory selection (N1 ERP component), (2) prevailing activation of parietal cortex of the right hemisphere (P4) at the stage of sensory categorization (P2 ERP component) as well as at the stage of the target stimulus identification (P3 ERP component). It is suggested that perception at the global level of the hierarchical stimulus is related primarily to the analysis of the spatial features of the stimulus in the dorsal visual system whereas the perception at the local level primarily involves an analysis of the object-related features in the ventral visual system.