Multiple Levels Of Representation Research Articles

Infant asphyxia detection using autoencoders trained on locally linear embedded-reduced Mel Frequency Cepstrum Coefficient (MFCC) features

Deep-learning methods are representation-learning methods with multiple levels of representation, obtained by composing simple but non-linear modules that each transform the representation at one level (starting with the raw input) into a representation at a higher, slightly more abstract level. With the composition of enough such transformations, very complex functions can be learned and solved. The objective of this paper is to explore one of the Deep Learning paradigms called autoencoders to perform diagnosis of infant asphyxia. LLE was used to reduce size of feature representation while enhancing them. Two stacked autoencoders were then trained to extract the necessary features for clasification. Extensive tests performed showed that the best-performing autoencoder network could produce 92.82% classification accuracy.

A quadratically regularized functional canonical correlation analysis for identifying the global structure of pleiotropy with NGS data.

Investigating the pleiotropic effects of genetic variants can increase statistical power, provide important information to achieve deep understanding of the complex genetic structures of disease, and offer powerful tools for designing effective treatments with fewer side effects. However, the current multiple phenotype association analysis paradigm lacks breadth (number of phenotypes and genetic variants jointly analyzed at the same time) and depth (hierarchical structure of phenotype and genotypes). A key issue for high dimensional pleiotropic analysis is to effectively extract informative internal representation and features from high dimensional genotype and phenotype data. To explore correlation information of genetic variants, effectively reduce data dimensions, and overcome critical barriers in advancing the development of novel statistical methods and computational algorithms for genetic pleiotropic analysis, we proposed a new statistic method referred to as a quadratically regularized functional CCA (QRFCCA) for association analysis which combines three approaches: (1) quadratically regularized matrix factorization, (2) functional data analysis and (3) canonical correlation analysis (CCA). Large-scale simulations show that the QRFCCA has a much higher power than that of the ten competing statistics while retaining the appropriate type 1 errors. To further evaluate performance, the QRFCCA and ten other statistics are applied to the whole genome sequencing dataset from the TwinsUK study. We identify a total of 79 genes with rare variants and 67 genes with common variants significantly associated with the 46 traits using QRFCCA. The results show that the QRFCCA substantially outperforms the ten other statistics.

Spatial pyramid face feature representation and weighted dissimilarity matching for improved face recognition

In this paper, we present a novel face recognition (FR) algorithm based on multiresolution spatial pyramid. In our method, a face is subdivided into increasingly finer subregions (local regions) and represented at multiple levels of histogram representations. To address image misalignment problem, overlapped patch-based local descriptor extraction has been also developed in an effective way. To preserve multiple levels of detail in facial local characteristics and to encode holistic spatial configuration, face features obtained for concatenated histograms (coming from all levels of spatial pyramid) are integrated into a combined feature set, termed spatial pyramid face feature representation (SPFR). In addition, to perform recognition by matching between the pair of probe and gallery SPFR sets, we propose the use of a weighted sum of the dissimilarity scores computed at all spatial pyramid levels. For this purpose, we develop a novel weight determination solution based on class-wise discriminant power estimation for face feature at a specific pyramid level. We incorporate our proposed algorithm into general FR pipeline and achieve encouraging identification results on the CMU-PIE, FERET, and LFW datasets, compared to previously developed methods. In addition, the feasibility of our method has been successfully demonstrated by making comparisons with other state-of-the-art FR methods (including deep CNN based method) under the FERET and FRGC 2.0 evaluation protocols. Based on results, our method is advantageous in terms of high recognition accuracy and low complexity, as well as straightforward implementation.

Model identification of reduced order fluid dynamics systems using deep learning

SummaryThis paper presents a novel model reduction method: deep learning reduced order model, which is based on proper orthogonal decomposition and deep learning methods. The deep learning approach is a recent technological advancement in the field of artificial neural networks. It has the advantage of learning the nonlinear system with multiple levels of representation and predicting data. In this work, the training data are obtained from high fidelity model solutions at selected time levels. The long short‐term memory network is used to construct a set of hypersurfaces representing the reduced fluid dynamic system. The model reduction method developed here is independent of the source code of the full physical system.The reduced order model based on deep learning has been implemented within an unstructured mesh finite element fluid model. The performance of the new reduced order model is evaluated using 2 numerical examples: an ocean gyre and flow past a cylinder. These results illustrate that the CPU cost is reduced by several orders of magnitude whilst providing reasonable accuracy in predictive numerical modelling.

Decoding levels of representation in reading: A representational similarity approach

Multiple levels of representation are involved in reading single words: visual representations of letter shape, orthographic representations of letter identity and order, phonological representations of the word's pronunciation, and semantic representations of its meaning. Previous lesion and neuroimaging studies have identified a network of regions recruited during word reading, including ventral occipital-temporal regions and the angular gyrus (AG). However, there is still debate about what information is being represented and processed in these regions. This study has two aims. The first is to help adjudicate between competing hypotheses concerning the role of ventral occipital cortex in reading. The second is to adjudicate between competing hypotheses concerning the role of the AG in reading. Participants read words in the scanner while performing a proper name detection task and we use a multivariate pattern analysis technique for analyzing fMRI data – representational similarity analysis (RSA) – to decode the type of information being represented in these regions based on computationally explicit theories. Distributed patterns of activation in the left ventral occipitotemporal cortex (vOT) and the AG show evidence of some type of orthographic processing, while the right hemisphere homologues of the vOT supports visual, but not orthographic, information processing of letter strings. In addition, there is evidence of left-lateralized semantic processing in the lvOT and evidence of top-down feedback in the lvOT. Taken together, these results suggest an interactive activation theory of visual word processing in which both the lvOT and lAG are neural loci of an orthographic level of representations.

Preparatory attention in visual cortex.

Top-down attention is the mechanism that allows us to selectively process goal-relevant aspects of a scene while ignoring irrelevant aspects. A large body of research has characterized the effects of attention on neural activity evoked by a visual stimulus. However, attention also includes a preparatory phase before stimulus onset in which the attended dimension is internally represented. Here, we review neurophysiological, functional magnetic resonance imaging, magnetoencephalography, electroencephalography, and transcranial magnetic stimulation (TMS) studies investigating the neural basis of preparatory attention, both when attention is directed to a location in space and when it is directed to nonspatial stimulus attributes (content-based attention) ranging from low-level features to object categories. Results show that both spatial and content-based attention lead to increased baseline activity in neural populations that selectively code for the attended attribute. TMS studies provide evidence that this preparatory activity is causally related to subsequent attentional selection and behavioral performance. Attention thus acts by preactivating selective neurons in the visual cortex before stimulus onset. This appears to be a general mechanism that can operate on multiple levels of representation. We discuss the functional relevance of this mechanism, its limitations, and its relation to working memory, imagery, and expectation. We conclude by outlining open questions and future directions.

Models for Music Analysis From a Markov Logic Networks Perspective

Analyzing and formalizing the intricate mechanisms of music is a very challenging goal for Artificial Intelligence. Dealing with real audio recordings requires the ability to handle both uncertainty and complex relational structure at multiple levels of representation. Until now, these two aspects have been generally treated separately, probability being the standard way to represent uncertainty in knowledge, while logical representation being the standard way to represent knowledge and complex relational information. Several approaches attempting a unification of logic and probability have recently been proposed. In particular, Markov logic networks MLNs, which combine first-order logic and probabilistic graphical models, have attracted increasing attention in recent years in many domains. This paper introduces MLNs as a highly flexible and expressive formalism for the analysis of music that encompasses most of the commonly used probabilistic and logic-based models. We first review and discuss existing approaches for music analysis. We then introduce MLNs in the context of music signal processing by providing a deep understanding of how they specifically relate to traditional models, specifically hidden Markov models and conditional random fields. We then present a detailed application of MLNs for tonal harmony music analysis that illustrates the potential of this framework for music processing.

Random Deep Belief Networks for Recognizing Emotions from Speech Signals.

Now the human emotions can be recognized from speech signals using machine learning methods; however, they are challenged by the lower recognition accuracies in real applications due to lack of the rich representation ability. Deep belief networks (DBN) can automatically discover the multiple levels of representations in speech signals. To make full of its advantages, this paper presents an ensemble of random deep belief networks (RDBN) method for speech emotion recognition. It firstly extracts the low level features of the input speech signal and then applies them to construct lots of random subspaces. Each random subspace is then provided for DBN to yield the higher level features as the input of the classifier to output an emotion label. All outputted emotion labels are then fused through the majority voting to decide the final emotion label for the input speech signal. The conducted experimental results on benchmark speech emotion databases show that RDBN has better accuracy than the compared methods for speech emotion recognition.

Syntactic levels, lexicalism, and ellipsis: The jury is still out.

Structural priming data are sometimes compatible with several theoretical views, as shown here for three key theoretical claims. One reason is that prime sentences affect multiple representational levels driving syntactic choice. Additionally, priming is affected by further cognitive functions (e.g., memory). We therefore see priming as a useful tool for the investigation of linguistic representation but not the only tool.

Developing knowledge of nonadjacent dependencies.

Characterizing the nature of linguistic representations and how they emerge during early development is a central goal in the cognitive science of language. One area in which this development plays out is in the acquisition of dependencies-relationships between co-occurring elements in a word, phrase, or sentence. These dependencies often involve multiple levels of representation and abstraction, built up as infants gain experience with their native language. The authors used the Headturn Preference Procedure to systematically investigate the early acquisition of 1 such dependency, the agreement between a subject and verb in French, at 6 different ages between 14 and 24 months. The results reveal a complex developmental trajectory that provides the first evidence that infants might indeed progress through distinct stages in the acquisition of this nonadjacent dependency. The authors discuss how changes in general cognition and representational knowledge (from reflecting surface statistics to higher-level morphological features) might account for their findings. These findings highlight the importance of studying language acquisition at close time intervals over a substantial age range. (PsycINFO Database Record

Event-related potential responses reveal simultaneous processing across multiple levels of representation in spoken word recognition

A controversial issue in spoken language comprehension concerns whether different sources of information are encapsulated from each other. Do listeners finish processing lower-level information (e.g., encoding acoustic differences) before beginning higher-level processing (e.g., determining the meaning of a word or its grammatical status)? We addressed these questions by examining the time-course of processing using an event-related potential experiment with a component-independent design. Listeners heard voiced/voiceless minimal pairs differing in (1) lexical status, (2) syntactic class (noun/verb distinctions), and (3) semantic content (animate/inanimate distinctions). For each voiced stimulus in a given condition (e.g., lexical status pair TUB/tup), there was a corresponding pair with a voiceless ending (tob/TOP). Stimuli were cross-spliced, allowing us to control for phonological and acoustic differences and examine higher-level effects independently of them. Widespread lexical status effects are observed shortly after the cross-splicing point (i.e., the time when the lexical properties of the word can first be determined) and persist for an extended time. Moreover, there is considerable overlap in the times during which both lexical status and semantic content effects are observed. These results suggest that listeners multiple types of linguistic representations are active simultaneously during spoken word recognition, consistent with parallel processing models.

Mining opinion summarizations using convolutional neural networks in Chinese microblogging systems

Chinese microblogging is an increasingly popular social media platform. Accurately summarizing representative opinions from microblogs can increase understanding of the semantics of opinions. The unique challenges of Chinese opinion summarization in microblogging systems are automatic learning of important features and selection of representative sentences. Deep-learning methods can automatically discover multiple levels of representations from raw data instead of requiring manual engineering. However, there have been very few systematic studies on sentiment analysis of Chinese hot topics using deep-learning methods. Based on the latest deep-learning research, in this paper, we propose a convolutional neural network (CNN)-based opinion summarization method for Chinese microblogging systems. The model first applies CNN to automatically mine useful features and perform sentiment analysis; then, by making good use of the obtained sentiment features, the semantic relationships among features are computed according to a hybrid ranking function; and finally, representative opinion sentences that are semantically related to the features are extracted using Maximal Marginal Relevance, which meets “relevant novelty” requirements. Experimental results on two real-world datasets verify the efficacy of the proposed model.

MapReduce based distributed learning algorithm for Restricted Boltzmann Machine

Deep learning is recently regarded as the closest artificial intelligence model to human brain. It is about learning multiple levels of representation and abstraction that help to make sense of data such as images, sound, and text. One deep model often consists of a hierarchical architecture that has the capability to model super non-linear and stochastic problems. Restricted Boltzmann Machine (RBM) is the main constructing block of current deep networks, as most of deep architectures are built with it. Based on MapReduce framework and Hadoop distributed file system, this paper proposes a distributed algorithm for training the RBM model. Its implementation and performance are evaluated on Big Data platform-Hadoop. The main contribution of the new learning algorithm is that it solves the scalability problem that limits the development of deep learning. The intelligence growing process of human brain requires learning from Big Data. The distributed learning mechanism for RBM makes it possible to abstract sophisticated and informative features from Big Data to achieve high-level intelligence. The evaluations of the proposed learning algorithm are carried out on image inpainting and classification problems based on the BAS dataset and MNIST hand-written digits dataset.

What do we mean by prediction in language comprehension?

ABSTRACTWe consider several key aspects of prediction in language comprehension: its computational nature, the representational level(s) at which we predict, whether we use higher-level representations to predictively pre-activate lower level representations, and whether we “commit” in any way to our predictions, beyond pre-activation. We argue that the bulk of behavioural and neural evidence suggests that we predict probabilistically and at multiple levels and grains of representation. We also argue that we can, in principle, use higher-level inferences to predictively pre-activate information at multiple lower representational levels. We suggest that the degree and level of predictive pre-activation might be a function of its expected utility, which, in turn, may depend on comprehenders’ goals and their estimates of the relative reliability of their prior knowledge and the bottom-up input. Finally, we argue that all these properties of language understanding can be naturally explained and productively explored within a multi-representational hierarchical actively generative architecture whose goal is to infer the message intended by the producer, and in which predictions play a crucial role in explaining the bottom-up input.

Re-examining selective adaptation: Fatiguing feature detectors, or distributional learning?

When a listener hears many good examples of a /b/ in a row, they are less likely to classify other sounds on, e.g., a /b/-to-/d/ continuum as /b/. This phenomenon is known as selective adaptation and is a well-studied property of speech perception. Traditionally, selective adaptation is seen as a mechanistic property of the speech perception system, and attributed to fatigue in acoustic-phonetic feature detectors. However, recent developments in our understanding of non-linguistic sensory adaptation and higher-level adaptive plasticity in speech perception and language comprehension suggest that it is time to re-visit the phenomenon of selective adaptation. We argue that selective adaptation is better thought of as a computational property of the speech perception system. Drawing on a common thread in recent work on both non-linguistic sensory adaptation and plasticity in language comprehension, we furthermore propose that selective adaptation can be seen as a consequence of distributional learning across multiple levels of representation. This proposal opens up new questions for research on selective adaptation itself, and also suggests that selective adaptation can be an important bridge between work on adaptation in low-level sensory systems and the complicated plasticity of the adult language comprehension system.

Neuroimaging the short- and long-term effects of repeated picture naming in healthy older adults

Repeated attempts to name pictures can improve subsequent naming for aphasic individuals with anomia, however, the neurocognitive mechanisms responsible for such improvements are unknown. This study investigated repeated picture naming in healthy older adults over a period of minutes (short-term) after one repetition and a period of days (long-term) after multiple repetitions. Compared to unprimed pictures, both repeated conditions showed faster naming latencies with the fastest latencies evident for the short-term condition. Neuroimaging results identified repetition suppression effects across three left inferior frontal gyrus regions of interest: for both the short- and long-term conditions in the pars orbitalis, and for long-term items in the pars triangularis and pars opercularis regions. The whole brain analysis also showed a repetition suppression effect in bilateral pars triangularis regions for the long-term condition. These findings within the inferior frontal gyrus suggest that effects of repeated naming may be driven by a mapping mechanism across multiple levels of representation, possibly reflecting different levels of learning, and lend support to the idea that processing may be hierarchically organised in the left inferior frontal gyrus. The whole brain analysis also revealed repetition suppression for the long-term condition within the posterior portion of bilateral inferior temporal gyri, which may reflect attenuation of integration processes within this region following the learning of task-relevant information.

Unrestrained beta reduction

A major argument for syntactic reconstruction is based on the well-known fact that semantic reconstruction by beta-reduction is possible only if the term to be substituted for a variable does not contain any variable that would become bound as a result of substitution. This way, we derive a theoretical argument for syntactic reconstruction. However, syntactic reconstruction is not without its problems, simply because the surface form and the reconstructed form may still differ with respect to other syntactic, semantic, and information theoretic properties. This is particularly troublesome for minimalist theories which do not allow for multiple levels of representation. In this paper we propose a technique that might help to overcome these difficulties (ie. the limitation imposed by beta-reduction on semantic reconstruction) by defining a translation function T for expressions of a predicate logic L0 with lambda-abstraction into expressions of a higher-order language L1, with the desirable property that the translation of a formula in L0 is equivalent with the translation of its unrestricted reduction. In linguistic applications this will facilitate the binding of a pronoun without presupposing c-command. We will sketch a formal proof showing that unrestricted beta-reduction is a property of the target expressions in L1, the translations of L0 under T.

Categorical biases in spatial memory: the role of certainty.

Memories for spatial locations often show systematic errors toward the central value of the surrounding region. The Category Adjustment (CA) model suggests that this bias is due to a Bayesian combination of categorical and metric information, which offers an optimal solution under conditions of uncertainty (Huttenlocher, Hedges, & Duncan, 1991). A fundamental assumption of this model is that representations of locations are unbiased but uncertain; during combination, greater metric uncertainty results in relatively greater emphasis on categorical information, ultimately leading to increased bias (but also minimizing error across multiple estimates). Sampaio and Wang (2009) have demonstrated that metric information is not lost during this combination process, supporting the CA model's assumption that underlying spatial representations are undistorted. Here, we examine the 2nd half of the CA model's central assumption: that increasing metric uncertainty drives the combination process. Participants recognized point locations within visually complex images in a 4-choice task. Our results indicate that individuals recognized the correct location over other, biased alternatives, confirming that metric information is unbiased at the time of retrieval. In addition, we found that, when participants make errors, they are more likely to select locations that are biased toward the category prototype. In Experiment 2, we demonstrate that categorically biased locations are most likely to be chosen under conditions of uncertainty. Indeed, under these conditions, categorically biased locations were chosen more frequently than the correct location. These results suggest that systematic errors are the result of combination across multiple levels of spatial representations that are undistorted but somewhat uncertain.

Music, neurology, and psychology in the nineteenth century.

This chapter examines connections between research in music, neurology, and psychology during the late-nineteenth century. Researchers in all three disciplines investigated how music is processed by the brain. Psychologists and comparative musicologists, such as Carl Stumpf, thought in terms of multiple levels of sensory processing and mental representation. Early thinking about music processing can be linked to the start of Gestalt psychology. Neurologists such as August Knoblauch also discussed multiple levels of music processing, basing speculation on ideas about language processing. Knoblauch and others attempted to localize music function in the brain. Other neurologists, such as John Hughlings Jackson, discussed a dissociation between music as an emotional system and language as an intellectual system. Richard Wallaschek seems to have been the only one from the late-nineteenth century to synthesize ideas from musicology, psychology, and neurology. He used ideas from psychology to explain music processing and audience reactions and also used case studies from neurology to support arguments about the nature of music. Understanding the history of this research sheds light on the development of all three disciplines-musicology, neurology, and psychology.

The influence of the visual modality on language structure and conventionalization: insights from sign language and gesture.

For humans, the ability to communicate and use language is instantiated not only in the vocal modality but also in the visual modality. The main examples of this are sign languages and (co-speech) gestures. Sign languages, the natural languages of Deaf communities, use systematic and conventionalized movements of the hands, face, and body for linguistic expression. Co-speech gestures, though non-linguistic, are produced in tight semantic and temporal integration with speech and constitute an integral part of language together with speech. The articles in this issue explore and document how gestures and sign languages are similar or different and how communicative expression in the visual modality can change from being gestural to grammatical in nature through processes of conventionalization. As such, this issue contributes to our understanding of how the visual modality shapes language and the emergence of linguistic structure in newly developing systems. Studying the relationship between signs and gestures provides a new window onto the human ability to recruit multiple levels of representation (e.g., categorical, gradient, iconic, abstract) in the service of using or creating conventionalized communicative systems.