Statistical Pattern Matching Research Articles

Disentangling Abstraction from Statistical Pattern Matching in Human and Machine Learning.

The ability to acquire abstract knowledge is a hallmark of human intelligence and is believed by many to be one of the core differences between humans and neural network models. Agents can be endowed with an inductive bias towards abstraction through meta-learning, where they are trained on a distribution of tasks that share some abstract structure that can be learned and applied. However, because neural networks are hard to interpret, it can be difficult to tell whether agents have learned the underlying abstraction, or alternatively statistical patterns that are characteristic of that abstraction. In this work, we compare the performance of humans and agents in a meta-reinforcement learning paradigm in which tasks are generated from abstract rules. We define a novel methodology for building "task metamers" that closely match the statistics of the abstract tasks but use a different underlying generative process, and evaluate performance on both abstract and metamer tasks. We find that humans perform better at abstract tasks than metamer tasks whereas common neural network architectures typically perform worse on the abstract tasks than the matched metamers. This work provides a foundation for characterizing differences between humans and machine learning that can be used in future work towards developing machines with more human-like behavior.

Visual Data Mining for Multimedia Databases

With computers and communication dominating technology in different fields, the need to look for media-based information processing, MBIP –rather than data-based information processing, DBIP- is increasingly being felt and this is compounded by the explosive developments in cellular communication, which brought computing and interaction on the move. The basis is to explore possibilities of using conventional data-mining approaches with visualization and object orientation so that human interaction is easier. Data Mining involves exploring databases to try and discover data relationships which are not explicitly stored with in the databases. Traditional techniques involve statistical analysis, clustering and pattern matching. Many current efforts are underway to integrate visualization in to this process. Visual data mining is a novel approach to data mining. The aim is to combine traditional data mining algorithms with information visualization techniques to utilize the advantages of both approaches. The utilization of both automatic analysis methods and human perception/understanding promises better and more effective data exploration. Visualization is a key process in visual data mining. Here the focus is on the presentation of all aspect of multimedia objects, their identification, their analysis and relationships.

Hybrid methodological approach to context-dependent speech recognition

Although the importance of contextual information in speech recognition has been acknowledged for a long time now, it has remained clearly underutilized even in state-of-the-art speech recognition systems. This article introduces a novel, methodologically hybrid approach to the research question of context-dependent speech recognition in human–machine interaction. To the extent that it is hybrid, the approach integrates aspects of both statistical and representational paradigms. We extend the standard statistical pattern-matching approach with a cognitively inspired and analytically tractable model with explanatory power. This methodological extension allows for accounting for contextual information which is otherwise unavailable in speech recognition systems, and using it to improve post-processing of recognition hypotheses. The article introduces an algorithm for evaluation of recognition hypotheses, illustrates it for concrete interaction domains, and discusses its implementation within two prototype conversational agents.

Analysis of area-specific expression patterns of RORbeta, ER81 and Nurr1 mRNAs in rat neocortex by double in situ hybridization and cortical box method.

BackgroundThe mammalian neocortex is subdivided into many areas, each of which exhibits distinctive lamina architecture. To investigate such area differences in detail, we chose three genes for comparative analyses, namely, RORbeta, ER81 and Nurr1, mRNAs of which have been reported to be mainly expressed in layers 4, 5 and 6, respectively. To analyze their qualitative and quantitative coexpression profiles in the rat neocortex, we used double in situ hybridization (ISH) histochemistry and cortical box method which we previously developed to integrate the data of different staining and individuals in a standard three-dimensional space.Principal FindingsOur new approach resulted in three main observations. First, the three genes showed unique area distribution patterns that are mostly complementary to one another. The patterns revealed by cortical box method matched well with the cytoarchitectonic areas defined by Nissl staining. Second, at single cell level, RORbeta and ER81 mRNAs were coexpressed in a subpopulation of layer 5 neurons, whereas Nurr1 and ER81 mRNAs were not colocalized. Third, principal component analysis showed that the order of hierarchical processing in the cortex correlates well with the expression profiles of these three genes. Based on this analysis, the dysgranular zone (DZ) in the somatosensory area was considered to exhibit a profile of a higher order area, which is consistent with previous proposal.Conclusions/SignificanceThe tight relationship between the expression of the three layer specific genes and functional areas were revealed, demonstrating the usefulness of cortical box method in the study on the cerebral cortex. In particular, it allowed us to perform statistical evaluation and pattern matching, which would become important in interpreting the ever-increasing data of gene expression in the cortex.

Statistical pattern matching facilitates the design of polyvalent inhibitors of anthrax and cholera toxins

Numerous biological processes involve the recognition of a specific pattern of binding sites on a target protein or surface. Although ligands displayed by disordered scaffolds form stochastic rather than specific patterns, theoretical models predict that recognition will occur between patterns that are characterized by similar or "matched" statistics. Endowing synthetic biomimetic structures with statistical pattern matching capabilities may improve the specificity of sensors and resolution of separation processes. We demonstrate that statistical pattern matching enhances the potency of polyvalent therapeutics. We functionalized liposomes with an inhibitory peptide at different densities and observed a transition in potency at an interpeptide separation that matches the distance between ligand-binding sites on the heptameric component of anthrax toxin. Pattern-matched polyvalent liposomes inhibited anthrax toxin in vitro at concentrations four orders of magnitude lower than the corresponding monovalent peptide, and neutralized this toxin in vivo. Statistical pattern matching also enhanced the potency of polyvalent inhibitors of cholera toxin. This facile strategy should be broadly applicable to the detection and neutralization of toxins and pathogens.

Acoustic backing-off as an implementation of missing feature theory

In this paper, we discuss acoustic backing-off as a method to improve automatic speech recognition robustness. Acoustic backing-off aims to achieve the same objective as the marginalization approach of missing feature theory: the detrimental influence of outlier values is effectively removed from the local distance computation in the Viterbi algorithm. The proposed method is based on one of the principles of robust statistical pattern matching: during recognition the local distance function (LDF) is modeled using a mixture of the distribution observed during training and a distribution describing observations not previously seen. In order to assess the effectiveness of the new method, we used artificial distortions of the acoustic vectors in connected digit recognition over telephone lines. We found that acoustic backing-off is capable of restoring recognition performance almost to the level observed for the undisturbed features, even in cases where a conventional LDF completely fails. These results show that recognition robustness can be improved using a marginalization approach, where making the distinction between reliable and corrupted feature values is wired into the recognition process. In addition, the results show that application of acoustic backing-off is not limited to feature representations based on filter bank outputs. Finally, the results indicate that acoustic backing-off is much less effective when local distortions are smeared over all vector elements. Therefore, the acoustic pre-processing steps should be chosen with care, so that the dispersion of distortions over all acoustic vector elements as a result of within-vector feature transformations is minimal.

Ticks and tick-borne disease systems in space and from space

Analyses within geographical information systems (GISs) indicate that small- and large-scale ranges of hard tick species (Ixodidae) are determined more by climate and vegetation than by host-related factors. Spatial distributions of ticks may therefore be analysed by statistical methods that seek correlations between known tick presence/absence and ground- or remotely-sensed (RS) environmental factors. In this way, local habitats of Amblyomma variegatum in the Caribbean and Ixodes ricinus in Europe have been mapped using Landsat RS imagery, while regional and continental distributions of African and temperate tick species have been predicted using multi-temporal information from the National Oceanic and Atmospheric Administration-Advanced Very High Resolution Radiometer (NOAA-AVHRR) imagery. These studies illustrate ways of maximizing statistical accuracy, whose interpretation is then discussed in a biological framework. Methods such as discriminant analysis are biologically transparent and interpretable, while others, such as logistic regression and tree-based classifications, are less so. Furthermore, the most consistently significant variable for predicting tick distributions, the RS Normalized Difference Vegetation Index (NDVI), has a sound biological basis in that it is related to moisture availability to free-living ticks and correlated with tick mortality rates. The development of biological process-based models for predicting the spatial dynamics of ticks is a top priority, especially as the risk of tick-borne infections is commonly related not simply to the vector's density, but to its seasonal population dynamics. Nevertheless, using statistical pattern-matching, the combination of RS temperature indices and NDVI successfully predicts certain temporal features essential for the transmission of tick-borne encephalitis virus, which translate into a spatial pattern of disease foci on a continental scale.

A simple theory and Monte Carlo simulations for recognition between random heteropolymers and disordered surfaces

We study the adsorption of random heteropolymers (RHPs) on disordered multifunctional surfaces. Recent replica calculations suggest that when the statistics describing the sequence distribution and the surface site distributions are related in a special way, a first-order adsorption transition occurs in such systems. This phenomenon of recognition between RHPs and disordered surfaces due to statistical pattern matching is studied via Monte Carlo simulations and a simple nonreplica theory. The theory serves to elucidate the simulation results, and suggests that the origin of the first-order adsorption transition is the suppression of loop fluctuations due to competing interactions and the quenched disorder (i.e., frustration).

Adsorption of random copolymers on disordered surfaces

Monte Carlo simulations are used to study adsorption of random heteropolymers on disordered substrates. The adsorbent surface is modelled as an impenetrable wall with the random distribution of sites characterised by short-ranged segment-specific interactions with the polymer. Spatial fluctuations in the interaction strength are described as a Gaussian process with zero mean and variance σ1 which is proportional to the site/segment potential, and to the loading of sites on the surface. In all cases, a moderate depletion in the polymer density is observed at small surface loading. Upon increasing σ1 the polymer begins to adsorb on the surface. At a certain threshold loading, we observe a sharp second adsorption transition followed by gradual saturation. The transition is interpreted in terms of pattern matching; i.e. the heteropolymers accomodate themselves at surface domains with site distributions that match the configuration and the sequence of the chain when the sequence and surface site distributions are related in a special way. Distinct peaks in heat capacity and compressibility observed at the adsorption transition point to a first order process envisaged in earlier replica/mean-field analysis. These results suggest that random heteropolymers with designed sequence statistics can ‘recognise’ multifunctional disordered surfaces due to statistical pattern-matching.

Check! A generic and specific industrial inspection tool

The authors present an overview of an 'inspection and control' package to perform automatic quality control of industrial components, specifically electronic circuit boards. This application has grown out of a set of tools designed initially to aid the development of point distribution models (PDMs) and grey-level models (GLMs) pioneered at the Wolfson Image Analysis Unit in Manchester. These models are both generic and specific. Generic, because they can be applied to most image classification problems; specific, because they aim at the full interpretation of the variability of the objects to be modelled. The underlying vision processing techniques are based on statistical pattern matching. Two synergetic approaches have been followed. The system is trained to recognise the components from either their shape or their grey level appearance or both.