Drop In Accuracy Research Articles

Classification of image matching point clouds over an urban area

ABSTRACTAirborne laser scanning (ALS) and image matching are the two main techniques for generating point clouds for large areas. While the classification of ALS point clouds has been well investigated, there are few studies that are related to image matching point clouds. In this study, point clouds of multiple resolutions from high-resolution aerial images (ground sampling distance, GSD, of 6 cm) over the city of Vienna were generated and investigated with respect to point density and processing time. Three different study sites with various urban structures are selected from a bigger dataset and classified based on two different approaches: machine learning and a traditional operator-based decision tree. Classification accuracy was evaluated and compared with confusion matrices. In general, the machine learning method results in a higher overall accuracy compared to the simple decision tree method, with accuracies of 87% and 84%, respectively, at the highest resolution. At lower-resolution levels (GSDs of 12 cm and 24 cm), the overall accuracy of machine learning drops by 4% and that of the simple decision tree by 7% for each level. Classifying rasterized data instead of the original point cloud resulted in an accuracy drop of 5%. Thus, using machine learning on point clouds at the highest available resolution is suggested for classification of urban areas.

Accuracy of Genomic Evaluations of Juvenile Growth Rate in Common Carp (Cyprinus carpio) Using Genotyping by Sequencing.

Cyprinids are the most important group of farmed fish globally in terms of production volume, with common carp (Cyprinus carpio) being one of the most valuable species of the group. The use of modern selective breeding methods in carp is at a formative stage, implying a large scope for genetic improvement of key production traits. In the current study, a population of 1,425 carp juveniles, originating from a partial factorial cross between 40 sires and 20 dams, was used for investigating the potential of genomic selection (GS) for juvenile growth, an exemplar polygenic production trait. RAD sequencing was used to identify and genotype SNP markers for subsequent parentage assignment, construction of a medium density genetic map (12,311 SNPs), genome-wide association study (GWAS), and testing of GS. A moderate heritability was estimated for body length of carp at 120 days (as a proxy of juvenile growth) of 0.33 (s.e. 0.05). No genome-wide significant QTL was identified using a single marker GWAS approach. Genomic prediction of breeding values outperformed pedigree-based prediction, resulting in 18% improvement in prediction accuracy. The impact of reduced SNP densities on prediction accuracy was tested by varying minor allele frequency (MAF) thresholds, with no drop in prediction accuracy until the MAF threshold is set <0.3 (2,744 SNPs). These results point to the potential for GS to improve economically important traits in common carp breeding programs.

Speaker recognition from whispered speech: A tutorial survey and an application of time-varying linear prediction

From the available biometric technologies, automatic speaker recognition is one of the most convenient and accessible ones due to abundance of mobile devices equipped with a microphone, allowing users to be authenticated across multiple environments and devices. Speaker recognition also finds use in forensics and surveillance. Due to the acoustic mismatch induced by varied environments and devices of the same speaker, leading to increased number of identification errors, much of the research focuses on compensating for such technology-induced variations, especially using machine learning at the statistical back-end. Another much less studied but at least as detrimental source of acoustic variation, however, arises from mismatched speaking styles induced by the speaker, leading to a substantial performance drop in recognition accuracy. This is a major problem especially in forensics where perpetrators may purposefully disguise their identity by varying their speaking style. We focus on one of the most commonly used ways of disguising one’s speaker identity, namely, whispering. We approach the problem of normal-whisper acoustic mismatch compensation from the viewpoint of robust feature extraction. Since whispered speech is intelligible, yet a low-intensity signal and therefore prone to extrinsic distortions, we take advantage of robust, long-term speech analysis methods that utilize slow articulatory movements in speech production. In specific, we address the problem using a novel method, frequency-domain linear prediction with time-varying linear prediction (FDLP-TVLP), which is an extension of the 2-dimensional autoregressive (2DAR) model that allows vocal tract filter parameters to be time-varying, rather than piecewise constant as in classic short-term speech analysis. Our speaker recognition experiments on the whisper subset of the CHAINS corpus indicate that when tested in normal-whisper mismatched conditions, the proposed FDLP-TVLP features improve speaker recognition performance by 7–10% over standard MFCC features in relative terms. We further observe that the proposed FDLP-TVLP features perform better than the FDLP and 2DAR methods for whispered speech.

Colorectal cancer detection by hyperspectral imaging using fluorescence excitation scanning.

Hyperspectral imaging technologies have shown great promise for biomedical applications. These techniques have been especially useful for detection of molecular events and characterization of cell, tissue, and biomaterial composition. Unfortunately, hyperspectral imaging technologies have been slow to translate to clinical devices - likely due to increased cost and complexity of the technology as well as long acquisition times often required to sample a spectral image. We have demonstrated that hyperspectral imaging approaches which scan the fluorescence excitation spectrum can provide increased signal strength and faster imaging, compared to traditional emission-scanning approaches. We have also demonstrated that excitation-scanning approaches may be able to detect spectral differences between colonic adenomas and adenocarcinomas and normal mucosa in flash-frozen tissues. Here, we report feasibility results from using excitation-scanning hyperspectral imaging to screen pairs of fresh tumoral and nontumoral colorectal tissues. Tissues were imaged using a novel hyperspectral imaging fluorescence excitation scanning microscope, sampling a wavelength range of 360-550 nm, at 5 nm increments. Image data were corrected to achieve a NIST-traceable flat spectral response. Image data were then analyzed using a range of supervised and unsupervised classification approaches within ENVI software (Harris Geospatial Solutions). Supervised classification resulted in >99% accuracy for single-patient image data, but only 64% accuracy for multi-patient classification (n=9 to date), with the drop in accuracy due to increased false-positive detection rates. Hence, initial data indicate that this approach may be a viable detection approach, but that larger patient sample sizes need to be evaluated and the effects of inter-patient variability studied.

A dynamic trust based two-layer neighbor selection scheme towards online recommender systems

Abstract Collaborative filtering has become one of the most widely used methods for providing recommendations in various online environments. Its recommendation accuracy highly relies on the selection of appropriate neighbors for the target user/item. However, existing neighbor selection schemes have some inevitable inadequacies, such as neglecting users’ capability of providing trustworthy recommendations, and ignoring users’ preference changes. Such inadequacies may lead to drop of the recommendation accuracy, especially when recommender systems are facing the data sparseness issue caused by the dramatic increase of users and items. To improve the recommendation accuracy, we propose a novel two-layer neighbor selection scheme that takes users’ capability and trustworthiness into account. In particular, the proposed scheme consists of two modules: (1) capability module that selects the first layer neighbors based on their capability of providing recommendations and (2) a trust module that further identifies the second layer neighbors based on their dynamic trustworthiness on recommendations. The performance of the proposed scheme is validated through experiments on real user datasets. Compared to three existing neighbor selection schemes, the proposed scheme consistently achieves the highest recommendation accuracy across data sets with different degrees of sparseness.

Second order finite-difference ghost-point multigrid methods for elliptic problems with discontinuous coefficients on an arbitrary interface

In this paper we propose a second-order accurate numerical method to solve elliptic problems with discontinuous coefficients (with general non-homogeneous jumps in the solution and its gradient) in 2D and 3D. The method consists of a finite-difference method on a Cartesian grid in which complex geometries (boundaries and interfaces) are embedded, and is second order accurate in the solution and the gradient itself. In order to avoid the drop in accuracy caused by the discontinuity of the coefficients across the interface, two numerical values are assigned on grid points that are close to the interface: a real value, that represents the numerical solution on that grid point, and a ghost value, that represents the numerical solution extrapolated from the other side of the interface, obtained by enforcing the assigned non-homogeneous jump conditions on the solution and its flux. The method is also extended to the case of matrix coefficient. The linear system arising from the discretization is solved by an efficient multigrid approach. Unlike the 1D case, grid points are not necessarily aligned with the normal derivative and therefore suitable stencils must be chosen to discretize interface conditions in order to achieve second order accuracy in the solution and its gradient. A proper treatment of the interface conditions will allow the multigrid to attain the optimal convergence factor, comparable with the one obtained by Local Fourier Analysis for rectangular domains. The method is robust enough to handle large jump in the coefficients: order of accuracy, monotonicity of the errors and good convergence factor are maintained by the scheme.

Fast Adjustable Threshold for Uniform Neural Network Quantization

Neural network quantization procedure is the necessary step for porting of neural networks to mobile devices. Quantization allows accelerating the inference, reducing memory consumption and model size. It can be performed without fine-tuning using calibration procedure (calculation of parameters necessary for quantization), or it is possible to train the network with quantization from scratch. Training with quantization from scratch on the labeled data is rather long and resource-consuming procedure. Quantization of network without fine-tuning leads to accuracy drop because of outliers which appear during the calibration. In this article we suggest to simplify the quantization procedure significantly by introducing the trained scale factors for quantization thresholds. It allows speeding up the process of quantization with fine-tuning up to 8 epochs as well as reducing the requirements to the set of train images. By our knowledge, the proposed method allowed us to get the first public available quantized version of MNAS without significant accuracy reduction - 74.8% vs 75.3% for original full-precision network. Model and code are ready for use and available at: this https URL.

From Winner-Takes-All to Winners-Share-All: Exploiting the Information Capacity in Temporal Codes.

In this letter, we have implemented and compared two neural coding algorithms in the networks of spiking neurons: Winner-takes-all (WTA) and winners-share-all (WSA). Winners-Share-All exploits the code space provided by the temporal code by training a different combination of [Formula: see text] out of [Formula: see text] neurons to fire together in response to different patterns, while WTA uses a one-hot-coding to respond to distinguished patterns. Using WSA, the maximum value of [Formula: see text] in order to maximize information capacity using [Formula: see text] output neurons was theoretically determined and utilized. A small proof-of-concept classification problem was applied to a spiking neural network using both algorithms to classify 14 letters of English alphabet with an image size of 15 [Formula: see text] 15 pixels. For both schemes, a modified spike-timing-dependent-plasticity (STDP) learning rule has been used to train the spiking neurons in an unsupervised fashion. The performance and the number of neurons required to perform this computation are compared between the two algorithms. We show that by tolerating a small drop in performance accuracy (84% in WSA versus 91% in WTA), we are able to reduce the number of output neurons by more than a factor of two. We show how the reduction in the number of neurons will increase as the number of patterns increases. The reduction in the number of output neurons would then proportionally reduce the number of training parameters, which requires less memory and hence speeds up the computation, and in the case of neuromorphic implementation on silicon, would take up much less area.

Impact of the face registration techniques on facial expressions recognition

Recent methodologies for facial expression recognition have been proposed and have obtained good results in near-frontal view. However, these situations do not fairly represent in-the-wild challenges, where expressions are natural and the subject is free of its movement. This is reflected in the accuracy drop of facial expression methods obtained on recent databases. Two challenges (head pose variations and large displacements) in facial expression recognition are studied in this paper. Experiments are proposed in order to quantify the impact of free head movements using representative expression recognition approaches (LBP, LBP-TOP, HOOF). We propose an experimental protocol (SNaP-2DFe) that records, under controlled light, facial expressions with two cameras: one attached on the head and one placed in front of the subject. As in both cameras facial expressions are the same, differences in performances measured on each camera show the impact of head pose variations and large displacements on the underlying recognition approach.

Polyploidism in Deep Neural Networks: m-Parent Evolutionary Synthesis of Deep Neural Networks in Varying Population Sizes

Evolutionary deep intelligence was recently proposed to organicallyproduce highly efficient deep neural network architecturesover successive generations. Thus far, current evolutionary synthesisprocesses are based on asexual reproduction, i.e., offspringneural networks are synthesized stochastically from a single parentnetwork. In this study, we investigate the effects of m-parentsexual evolutionary synthesis (m = 1, 2, 3, 5) in combination withvarying population sizes of three, five, and eight synthesized networksper generation. Experimental results were obtained usinga 10% subset of the MNIST handwritten digits dataset, and showthat increasing the number of parent networks results in improvedarchitectural efficiency of the synthesized networks (approximately150x synaptic efficiency and approximately 42–49x cluster efficiency)while resulting in only a 2–3% drop in testing accuracy.

Irrelevant information in math problems need not be inhibited: Students might just need to spot them

This study examined whether individual differences in inhibitory abilities were related to word problem-solving performance. A sample of 10–11year-old students (N=134) were assessed on two types of inhibition: prepotent response inhibition and resistance to proactive interference. Word problems administered contained varying amounts of either numerical or literal irrelevant information. Working memory capacity and the students' ability to identify irrelevant information were also assessed. Numerical but not literal irrelevant information resulted in poorer problem-solving accuracy. Relevancy identification was associated with the observed drop in accuracy when numerical irrelevant information was added to word problems. Furthermore, neither inhibitory skills nor working memory explained significance variance in accuracy drop. We discussed these findings in relation to other research that considered inhibitory abilities and word problem-solving. We also discussed students' problem-solving errors in terms of strategy use and, in the larger socio-mathematical context.

A cascading scheme for speeding up multiple classifier systems

The accuracy of multi-class classification problems is improving at a good pace. However, improving the accuracy often leads to slowing down the processing speed. Since employing a large number of classifiers or a combination of them is a time-consuming process, the sluggish behavior is more evident in multiple classifier systems. In this paper, a practical cascading scheme is proposed for boosting the speed of a multiple binary classifier system with no noticeable reduction in recognition rate. The proposed cascading scheme sets a sequence of binary classifiers and applies one classifier at a time to the input. Some effective criteria for a practical ordering of classifiers are introduced, and a fusion of them is verified to be the best. A vehicle make and model recognition (VMMR) system with multiple individual classifiers is presented briefly as a use case for a multi-class classification problem. The experiments done on this VMMR system using two completely different datasets confirm the effectiveness of our scheme. One of the configurations of the proposed scheme results in up to 30% speedup in comparison with the baseline VMMR system with analogous recognition rate. Another configuration of the proposed cascading scheme achieves up to 80% speedup with just a minor drop in accuracy.

A Tucker Deep Computation Model for Mobile Multimedia Feature Learning

Recently, the deep computation model, as a tensor deep learning model, has achieved super performance for multimedia feature learning. However, the conventional deep computation model involves a large number of parameters. Typically, training a deep computation model with millions of parameters needs high-performance servers with large-scale memory and powerful computing units, limiting the growth of the model size for multimedia feature learning on common devices such as portable CPUs and conventional desktops. To tackle this problem, this article proposes a Tucker deep computation model by using the Tucker decomposition to compress the weight tensors in the full-connected layers for multimedia feature learning. Furthermore, a learning algorithm based on the back-propagation strategy is devised to train the parameters of the Tucker deep computation model. Finally, the performance of the Tucker deep computation model is evaluated by comparing with the conventional deep computation model on two representative multimedia datasets, that is, CUAVE and SNAE2, in terms of accuracy drop, parameter reduction, and speedup in the experiments. Results imply that the Tucker deep computation model can achieve a large-parameter reduction and speedup with a small accuracy drop for multimedia feature learning.

To use or not to use: Feature selection for sentiment analysis of highly imbalanced data

AbstractWe investigate feature selection methods for machine learning approaches in sentiment analysis. More specifically, we use data from the cooking platform Epicurious and attempt to predict ratings for recipes based on user reviews. In machine learning approaches to such tasks, it is a common approach to use word or part-of-speech n-grams. This results in a large set of features, out of which only a small subset may be good indicators for the sentiment. One of the questions we investigate concerns the extension of feature selection methods from a binary classification setting to a multi-class problem. We show that an inherently multi-class approach, multi-class information gain, outperforms ensembles of binary methods. We also investigate how to mitigate the effects of extreme skewing in our data set by making our features more robust and by using review and recipe sampling. We show that over-sampling is the best method for boosting performance on the minority classes, but it also results in a severe drop in overall accuracy of at least 6 per cent points.

Learning from a Class Imbalanced Public Health Dataset: a Cost-based Comparison of Classifier Performance

Public health care systems routinely collect health-related data from the population. This data can be analyzed using data mining techniques to find novel, interesting patterns, which could help formulate effective public health policies and interventions. The occurrence of chronic illness is rare in the population and the effect of this class imbalance, on the performance of various classifiers was studied. The objective of this work is to identify the best classifiers for class imbalanced health datasets through a cost-based comparison of classifier performance. The popular, open-source data mining tool WEKA, was used to build a variety of core classifiers as well as classifier ensembles, to evaluate the classifiers’ performance. The unequal misclassification costs were represented in a cost matrix, and cost-benefit analysis was also performed. In another experiment, various sampling methods such as under-sampling, over-sampling, and SMOTE was performed to balance the class distribution in the dataset, and the costs were compared. The Bayesian classifiers performed well with a high recall, low number of false negatives and were not affected by the class imbalance. Results confirm that total cost of Bayesian classifiers can be further reduced using cost-sensitive learning methods. Classifiers built using the random under-sampled dataset showed a dramatic drop in costs and high classification accuracy.

Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition

Like other problems in computer vision, offline handwritten Chinese character recognition (HCCR) has achieved impressive results using convolutional neural network (CNN)-based methods. However, larger and deeper networks are needed to deliver state-of-the-art results in this domain. Such networks intuitively appear to incur high computational cost, and require the storage of a large number of parameters, which render them unfeasible for deployment in portable devices. To solve this problem, we propose a Global Supervised Low-rank Expansion (GSLRE) method and an Adaptive Drop-weight (ADW) technique to solve the problems of speed and storage capacity. We design a nine-layer CNN for HCCR consisting of 3755 classes, and devise an algorithm that can reduce the network’s computational cost by nine times and compress the network to 1/18 of the original size of the baseline model, with only a 0.21% drop in accuracy. In tests, the proposed algorithm can still surpass the best single-network performance reported thus far in the literature while requiring only 2.3MB for storage. Furthermore, when integrated with our effective forward implementation, the recognition of an offline character image takes only 9.7 ms on a CPU. Compared with the state-of-the-art CNN model for HCCR, our approach is approximately 30 times faster, yet 10 times more cost efficient.

The Sustained Influence of an Error on Future Decision-Making.

Post-error slowing (PES) is consistently observed in decision-making tasks after negative feedback. Yet, findings are inconclusive as to whether PES supports performance accuracy. We addressed the role of PES by employing drift diffusion modeling which enabled us to investigate latent processes of reaction times and accuracy on a large-scale dataset (>5,800 participants) of a visual search experiment with emotional face stimuli. In our experiment, post-error trials were characterized by both adaptive and non-adaptive decision processes. An adaptive increase in participants’ response threshold was sustained over several trials post-error. Contrarily, an initial decrease in evidence accumulation rate, followed by an increase on the subsequent trials, indicates a momentary distraction of task-relevant attention and resulted in an initial accuracy drop. Higher values of decision threshold and evidence accumulation on the post-error trial were associated with higher accuracy on subsequent trials which further gives credence to these parameters’ role in post-error adaptation. Finally, the evidence accumulation rate post-error decreased when the error trial presented angry faces, a finding suggesting that the post-error decision can be influenced by the error context. In conclusion, we demonstrate that error-related response adaptations are multi-component processes that change dynamically over several trials post-error.

Atrial fibrillation detection on compressed sensed ECG

Objective: Compressive sensing (CS) approaches to electrocardiogram (ECG) analysis provide efficient methods for real time encoding of cardiac activity. In doing so, it is important to assess the downstream effect of the compression on any signal processing and classification algorithms. CS is particularly suitable for low power wearable devices, thanks to its low-complex digital or hardware implementation that directly acquires a compressed version of the signal through random projections. In this work, we evaluate the impact of CS compression on atrial fibrillation (AF) detection accuracy. Approach: We compare schemes with data reconstruction based on wavelet and Gaussian models, followed by a P&T-based identification of beat-to-beat (RR) intervals on the MIT-BIH atrial fibrillation database. A state-of-the-art AF detector is applied to the RR time series and the accuracy of the AF detector is then evaluated under different levels of compression. We also consider a new beat detection procedure which operates directly in the compressed domain, avoiding costly signal reconstruction procedures. Main results: We demonstrate that for compression ratios up to 30 the AF detector applied to RR intervals derived from the compressed signal exhibits results comparable to those achieved when employing a standard QRS detector on the raw uncompressed signals, and exhibits only a 2% accuracy drop at a compression ratio of 60%. We also show that the Gaussian-based reconstruction approach is superior in terms of AF detection accuracy, with a negligible drop in performance at a compression ratio ⩽75%, compared to a wavelet approach, which exhibited a significant drop in accuracy at a compression ratio ⩾65%. Significance: The results suggest that CS should be considered as a plausible methodology for both efficient real time ECG compression (at moderate compression rates) and for offline analysis (at high compression rates).

Pan‐lymphoma classification

Introduction: Cell of origin classification of diffuse large B-cell lymphoma (DL-BCL) using gene expression signatures identifies patients with distinct molecular features and survival outcomes. Whilst the identification of activated B-cell (ABC) and germinal center B-cell (GCB) subtypes of DL-BCL is now routinely used in diagnostics, however, a significant proportion of patients display gene expression signatures that are intermediate between classes. Whilst some of these intermediate cases display genuine hallmarks of both ABC or GCB subtypes, leading to an intermediate DL-BCL COO-UNC type, others appear to display distinct molecular signatures. It is possible that that these signatures represent inherent noise in the expression profiling process, or biological noise resulting from, for example, high levels of T-cell infiltration in samples. A more intriguing possibility is that these cases represent intermediate types between DL-BCL and other types of lymphoma. The existence of intermediate molecular signatures raises important questions about which type of treatment is most applicable for these cases. Yet the identification of intermediate molecular signatures is challenging, because existing classification algorithms have tended to focus on a small subset of lymphoma types rather than across the pan-lymphoma spectrum. Moreover, it is not clear what features would be needed to identify the relationship between patients to diverse lymphoma classes. Methods: To address this issue, we have developed a pan-lymphoma classifier using a support vector machine (SVM) with an embedded recursive feature selection (RFS) algorithm. Our training and test dataset consisted of 431 samples, spanning eight different types of lymphoma: ABC (49), GCB (133), COO-UNC (40), Burkitt's (60), Hodgkin's (43), PMBL (35), PBL (39), and Plasmacytoma (32). Gene expression profiles for each sample were obtained using the Illumina DASL platform. We trained the SVM using 10-fold cross-validation with a 70:30 train/test split. On each iteration of the RFS, we removed the single least informative probe. We repeated this until a significant drop in average accuracy of the SVM was reached. Results: Initially training the SVM on 1125 probes, we obtained a pan-lymphoma average classification accuracy of 91%. We then used the SVM-RFS to obtain a list of just 23 probes that gave rise to an average classification accuracy of 86%. This allowed us to develop a between-class distance metric as a score of the relative association of any given sample to a lymphoma class. Retrospective classification of COO-UNC samples showed some share molecular features associated with non-DL-BCL classes. Conclusions: SVM-RFS appears to be a robust approach for identifying classes of lymphoma from underlying molecular phenotypes. A small number of most informative genes provide a separation of cases into subtypes that is in excellent agreement with laboratory diagnoses. Keywords: activated B-cell–like (ABC); GCB lymphoma subtype; gene expression profile (GEP)

Generating Up-to-Date and Detailed Land Use and Land Cover Maps Using OpenStreetMap and GlobeLand30

With the opening up of the Landsat archive, global high resolution land cover maps have begun to appear. However, they often have only a small number of high level land cover classes and they are static products, corresponding to a particular period of time, e.g., the GlobeLand30 (GL30) map for 2010. The OpenStreetMap (OSM), in contrast, consists of a very detailed, dynamically updated, spatial database of mapped features from around the world, but it suffers from incomplete coverage, and layers of overlapping features that are tagged in a variety of ways. However, it clearly has potential for land use and land cover (LULC) mapping. Thus the aim of this paper is to demonstrate how the OSM can be converted into a LULC map and how this OSM-derived LULC map can then be used to first update the GL30 with more recent information and secondly, enhance the information content of the classes. The technique is demonstrated on two study areas where there is availability of OSM data but in locations where authoritative data are lacking, i.e., Kathmandu, Nepal and Dar es Salaam, Tanzania. The GL30 and its updated and enhanced versions are independently validated using a stratified random sample so that the three maps can be compared. The results show that the updated version of GL30 improves in terms of overall accuracy since certain classes were not captured well in the original GL30 (e.g., water in Kathmandu and water/wetlands in Dar es Salaam). In contrast, the enhanced GL30, which contains more detailed urban classes, results in a drop in the overall accuracy, possibly due to the increased number of classes, but the advantages include the appearance of more detailed features, such as the road network, that becomes clearly visible.