Recent Deep Neural Networks Research Articles

Have We Solved Edge Detection? A Review of State-of-the-Art Datasets and DNN Based Techniques

Recent Deep Neural Networks (DNNs) based edge detection methods claim to beat human performance on small scale datasets like BSDS500. But is the problem of edge detection really solved? To answer this question, we review the existing dataset limitations as well as the generalization capabilities of the proposed architectures. To this end, we develop a Synthetic Textured Masks Dataset (STMD) that contains 28,000 gray scale images. The performance of several edge detection methods is severely degraded on STMD. To further validate these results we propose a baseline Single Scale Feed Forward Edge Detector (SFED), which is a simple 9-layer feed-forward convolutional neural network with no pooling layers. The performance of SFED is better than most state-of-the-art architectures on BSDS500 and is superior to all the compared architectures on STMD. These results show that most of the architectural advancements of existing architectures are at the cost of generalizability where if we change the dataset set distribution (both training and testset), the performance become significantly degraded and therefore the problem of edge detection is still far away from being solved. There are also severe limitations of existing datasets in the field, and STMD provides a framework for designing and testing better edge detection architectures for novel application areas, such as, medical imaging and self-driving cars.

Meta PID Attention Network for Flexible and Efficient Real-World Noisy Image Denoising.

Recent deep convolutional neural networks for real-world noisy image denoising have shown a huge boost in performance by training a well-engineered network over external image pairs. However, most of these methods are generally trained with supervision. Once the testing data is no longer compatible with the training conditions, they can exhibit poor generalization and easily result in severe overfitting or degrading performances. To tackle this barrier, we propose a novel denoising algorithm, dubbed as Meta PID Attention Network (MPA-Net). Our MPA-Net is built based upon stacking Meta PID Attention Modules (MPAMs). In each MPAM, we utilize a second-order attention module (SAM) to exploit the channel-wise feature correlations with second-order statistics, which are then adaptively updated via a proportional-integral-derivative (PID) guided meta-learning framework. This learning framework exerts the unique property of the PID controller and meta-learning scheme to dynamically generate filter weights for beneficial update of the extracted features within a feedback control system. Moreover, the dynamic nature of the framework enables the generated weights to be flexibly tweaked according to the input at test time. Thus, MPAM not only achieves discriminative feature learning, but also facilitates a robust generalization ability on distinct noises for real images. Extensive experiments on ten datasets are conducted to inspect the effectiveness of the proposed MPA-Net quantitatively and qualitatively, which demonstrates both its superior denoising performance and promising generalization ability that goes beyond those of the state-of-the-art denoising methods.

A Comprehensive Study on Deep Learning-Based Methods for Sign Language Recognition

In this paper, a comparative experimental assessment of computer vision-based methods for sign language recognition is conducted. By implementing the most recent deep neural network methods in this field, a thorough evaluation on multiple publicly available datasets is performed. The aim of the present study is to provide insights on sign language recognition, focusing on mapping non-segmented video streams to glosses. For this task, two new sequence training criteria, known from the fields of speech and scene text recognition, are introduced. Furthermore, a plethora of pretraining schemes is thoroughly discussed. Finally, a new RGB+D dataset for the Greek sign language is created. To the best of our knowledge, this is the first sign language dataset where sentence and gloss level annotations are provided for a video capture.

Adversarial Training for Multi Domain Dialog System

Natural Language Understanding and Speech Understanding systems are now a global trend, and with the advancement of artificial intelligence and machine learning techniques, have drawn attention from both the academic and business communities. Domain prediction, intent detection and entity extraction or slot fillings are the most important parts for such intelligent systems. Various traditional machine learning algorithms such as Bayesian algorithm, Support Vector Machine, and Artificial Neural Network, along with recent Deep Neural Network techniques, are used to predict domain, intent, and entity. Most language understanding systems process user input in a sequential order: domain is first predicted, then intent and slots are filled according to the semantic frames of the predicted domain. This pipeline approach, however, has many disadvantages including downstream error; i.e., if the system fails to predict the domain, then the system also fails to predict intent and slot. The main purpose of this paper is to mitigate the risk of downstream error propagation in traditional pipelined models and improve the predictive performance of domain, intent, and slot- all of which are critical steps for speech understanding and dialog systems- with a deep learning-based single joint model trained with an adversarial approach and long short-term memory (LSTM) algorithm. The systematic experimental analysis shows significant improvements in predictive performance for domain, intent, and entity with the proposed adversarial joint model, compared to the base joint model.

Syntax Role for Neural Semantic Role Labeling

Semantic role labeling (SRL) is dedicated to recognizing the semantic predicate-argument structure of a sentence. Previous studies in terms of traditional models have shown syntactic information can make remarkable contributions to SRL performance; however, the necessity of syntactic information was challenged by a few recent neural SRL studies that demonstrate impressive performance without syntactic backbones and suggest that syntax information becomes much less important for neural semantic role labeling, especially when paired with recent deep neural network and large-scale pre-trained language models. Despite this notion, the neural SRL field still lacks a systematic and full investigation on the relevance of syntactic information in SRL, for both dependency and both monolingual and multilingual settings. This paper intends to quantify the importance of syntactic information for neural SRL in the deep learning framework. We introduce three typical SRL frameworks (baselines), sequence-based, tree-based, and graph-based, which are accompanied by two categories of exploiting syntactic information: syntax pruning-based and syntax feature-based. Experiments are conducted on the CoNLL-2005, -2009, and -2012 benchmarks for all languages available, and results show that neural SRL models can still benefit from syntactic information under certain conditions. Furthermore, we show the quantitative significance of syntax to neural SRL models together with a thorough empirical survey using existing models.

Automated Detection of COVID-19 Cases using Recent Deep Convolutional Neural Networks and CT images.

COVID-19 is an acute severe respiratory disease caused by a novel coronavirus SARS-CoV-2. After its first appearance in Wuhan (China), it spread rapidly across the world and became a pandemic. It had a devastating effect on everyday life, public health, and the world economy. The use of advanced artificial intelligence (AI) techniques combined with radiological imaging can be helpful in speeding-up the detection of this disease. In this study, we propose the development of recent deep learning models for automatic COVID-19 detection using computed tomography (CT) images. The proposed models are fine-tuned and optimized to provide accurate results for multiclass classification of COVID-19 vs. Community Acquired Pneumonia (CAP) vs. Normal cases. Tests were conducted both at the image and patient-level and show that the proposed algorithms achieve very high scores. In addition, an explainability algorithm was developed to help visualize the symptoms of the disease detected by the best performing deep model.

Learning-Rate Annealing Methods for Deep Neural Networks

Deep neural networks (DNNs) have achieved great success in the last decades. DNN is optimized using the stochastic gradient descent (SGD) with learning rate annealing that overtakes the adaptive methods in many tasks. However, there is no common choice regarding the scheduled-annealing for SGD. This paper aims to present empirical analysis of learning rate annealing based on the experimental results using the major data-sets on the image classification that is one of the key applications of the DNNs. Our experiment involves recent deep neural network models in combination with a variety of learning rate annealing methods. We also propose an annealing combining the sigmoid function with warmup that is shown to overtake both the adaptive methods and the other existing schedules in accuracy in most cases with DNNs.

Detection of tuberculosis from chest X-ray images: Boosting the performance with vision transformer and transfer learning

Tuberculosis (TB) caused by Mycobacterium tuberculosis is a contagious disease which is among the top deadly diseases in the world. Research in Medical Imaging has been done to provide doctors with techniques and tools to early detect, monitor and diagnose the disease using Artificial Intelligence. Recently, many attempts have been made to automatically recognize TB from chest X-ray (CXR) images. Still, while the obtained performance is encouraging, according to our investigation, many of the existing approaches have been evaluated on small and undiverse datasets. We suppose that such a good performance might not hold for heterogeneous data sources, which originate from real world scenarios. Our present work aims to fill the gap and improve the prediction performance on larger datasets. In particular, we present a practical solution for the detection of tuberculosis from CXR images, making use of cutting-edge Machine Learning and Computer Vision algorithms. We conceptualize a framework by adopting three recent deep neural networks as the main classification engines, namely modified EfficientNet, modified original Vision Transformer, and modified Hybrid EfficientNet with Vision Transformer. Moreover, we also empower the learning process with various augmentation techniques. We evaluated the proposed approach using a large dataset which has been curated by merging various public datasets. The resulting dataset has been split into training, validation, and testing sets which account for 80%, 10%, and 10% of the original dataset, respectively. To further study our proposed approach, we compared it with two state-of-the-art systems. The obtained results are encouraging: the maximum accuracy of 97.72% with AUC of 100% is achieved with ViT_Base_EfficientNet_B1_224. The experimental results demonstrate that our conceived tool outperforms the considered baselines with respect to different quality metrics.

Text‐based question answering from information retrieval and deep neural network perspectives: A survey

AbstractText‐based question answering (QA) is a challenging task which aims at finding short concrete answers for users' questions. This line of research has been widely studied with information retrieval (IR) techniques and has received increasing attention in recent years by considering deep neural network approaches. Deep learning (DL) approaches, which are the main focus of this paper, provide a powerful technique to learn multiple layers of representations and interaction between the questions and the answer sentences. In this paper, we provide a comprehensive overview of different models proposed for the QA task, including both a traditional IR perspective and a more recent deep neural network environment. We also introduce well‐known datasets for the task and present available results from the literature to have a comparison between different techniques.This article is categorized under: Algorithmic Development > Text Mining Technologies > Machine Learning

Learning to Purify Noisy Labels via Meta Soft Label Corrector

Recent deep neural networks (DNNs) can easily overfit to biased training data with noisy labels. Label correction strategy is commonly used to alleviate this issue by identifying suspected noisy labels and then correcting them. Current approaches to correcting corrupted labels usually need manually pre-defined label correction rules, which makes it hard to apply in practice due to the large variations of such manual strategies with respect to different problems. To address this issue, we propose a meta-learning model, aiming at attaining an automatic scheme which can estimate soft labels through meta-gradient descent step under the guidance of a small amount of noise-free meta data. By viewing the label correction procedure as a meta-process and using a meta-learner to automatically correct labels, our method can adaptively obtain rectified soft labels gradually in iteration according to current training problems. Besides, our method is model-agnostic and can be combined with any other existing classification models with ease to make it available to noisy label cases. Comprehensive experiments substantiate the superiority of our method in both synthetic and real-world problems with noisy labels compared with current state-of-the-art label correction strategies.

Deep multilabel CNN for forensic footwear impression descriptor identification

In recent years deep neural networks have become the workhorse of computer vision. In this paper, we employ a deep learning approach to classify footwear impression’s features known as descriptors for forensic use cases. Within this process, we develop and evaluate an effective technique for feeding downsampled greyscale impressions to a neural network pre-trained on data from a different domain. Our approach relies on learnable preprocessing layer paired with multiple interpolation methods used in parallel. We empirically show that this technique outperforms using a single type of interpolated image without learnable preprocessing, and can help to avoid the computational penalty related to using high resolution inputs, by making more efficient use of the low resolution inputs. We also investigate the effect of preserving the aspect ratio of the inputs, which leads to considerable boost in accuracy without increasing the computational budget with respect to squished rectangular images. Finally, we formulate a set of best practices for transfer learning with greyscale inputs, potentially widely applicable in computer vision tasks ranging from footwear impression classification to medical imaging.

Compressive spectral image reconstruction using deep prior and low-rank tensor representation.

Compressive spectral imaging (CSI) has emerged as an alternative spectral image acquisition technology, which reduces the number of measurements at the cost of requiring a recovery process. In general, the reconstruction methods are based on handcrafted priors used as regularizers in optimization algorithms or recent deep neural networks employed as an image generator to learn a non-linear mapping from the low-dimensional compressed measurements to the image space. However, these deep learning methods need many spectral images to obtain good performance. In this work, a deep recovery framework for CSI without training data is presented. The proposed method is based on the fact that the structure of some deep neural networks and an appropriated low-dimensional structure are sufficient to impose a structure of the underlying spectral image from CSI. We analyzed the low-dimensional structure via the Tucker representation, modeled in the first net layer. The proposed scheme is obtained by minimizing the ${\ell _2}$-norm distance between the compressive measurements and the predicted measurements, and the desired recovered spectral image is formed just before the forward operator. Simulated and experimental results verify the effectiveness of the proposed method for the coded aperture snapshot spectral imaging.

Representational Content of Oscillatory Brain Activity during Object Recognition: Contrasting Cortical and Deep Neural Network Hierarchies.

Numerous theories propose a key role for brain oscillations in visual perception. Most of these theories postulate that sensory information is encoded in specific oscillatory components (e.g., power or phase) of specific frequency bands. These theories are often tested with whole-brain recording methods of low spatial resolution (EEG or MEG), or depth recordings that provide a local, incomplete view of the brain. Opportunities to bridge the gap between local neural populations and whole-brain signals are rare. Here, using representational similarity analysis (RSA) in human participants we explore which MEG oscillatory components (power and phase, across various frequency bands) correspond to low or high-level visual object representations, using brain representations from fMRI, or layer-wise representations in seven recent deep neural networks (DNNs), as a template for low/high-level object representations. The results showed that around stimulus onset and offset, most transient oscillatory signals correlated with low-level brain patterns (V1). During stimulus presentation, sustained β (∼20 Hz) and γ (>60 Hz) power best correlated with V1, while oscillatory phase components correlated with IT representations. Surprisingly, this pattern of results did not always correspond to low-level or high-level DNN layer activity. In particular, sustained β band oscillatory power reflected high-level DNN layers, suggestive of a feed-back component. These results begin to bridge the gap between whole-brain oscillatory signals and object representations supported by local neuronal activations.

The whole is\xa0greater than\xa0its\xa0parts: ensembling improves protein contact prediction

The prediction of amino acid contacts from protein sequence is an important problem, as protein contacts are a vital step towards the prediction of folded protein structures. We propose that a powerful concept from deep learning, called ensembling, can increase the accuracy of protein contact predictions by combining the outputs of different neural network models. We show that ensembling the predictions made by different groups at the recent Critical Assessment of Protein Structure Prediction (CASP13) outperforms all individual groups. Further, we show that contacts derived from the distance predictions of three additional deep neural networks—AlphaFold, trRosetta, and ProSPr—can be substantially improved by ensembling all three networks. We also show that ensembling these recent deep neural networks with the best CASP13 group creates a superior contact prediction tool. Finally, we demonstrate that two ensembled networks can successfully differentiate between the folds of two highly homologous sequences. In order to build further on these findings, we propose the creation of a better protein contact benchmark set and additional open-source contact prediction methods.

Sketch-specific data augmentation for freehand sketch recognition

Sketch recognition remains a significant challenge due to the limited training data and the substantial intra-class variance of freehand sketches for the same object. Conventional methods for this task often rely on the availability of the temporal order of sketch strokes, additional cues acquired from different modalities and supervised augmentation of sketch datasets with real images, which also limit the applicability and feasibility of these methods in real scenarios.In this paper, we propose a novel sketch-specific data augmentation (SSDA) method that leverages the quantity and quality of the sketches automatically. From the aspect of quantity, we introduce a Bezier pivot based deformation (BPD) strategy to enrich the training data. Towards quality improvement, we present a mean stroke reconstruction (MSR) approach to generate a set of novel types of sketches with smaller intra-class variances. Both of these solutions are unrestricted from any multi-source data and temporal cues of sketches. Furthermore, we show that some recent deep convolutional neural network models that are trained on generic classes of real images can be better choices than most of the elaborate architectures that are designed explicitly for sketch recognition. As SSDA can be integrated with any convolutional neural networks, it has a distinct advantage over the existing methods. Our extensive experimental evaluations demonstrate that the proposed method achieves the state-of-the-art results (84.27%) on the TU-Berlin dataset, outperforming the human performance by a remarkable 11.17% increase. Finally, more experiments show the practical value of our approach for the task of sketch-based image retrieval.

Rare bioparticle detection via deep metric learning.

Recent deep neural networks have shown superb performance in analyzing bioimages for disease diagnosis and bioparticle classification. Conventional deep neural networks use simple classifiers such as SoftMax to obtain highly accurate results. However, they have limitations in many practical applications that require both low false alarm rate and high recovery rate, e.g., rare bioparticle detection, in which the representative image data is hard to collect, the training data is imbalanced, and the input images in inference time could be different from the training images. Deep metric learning offers a better generatability by using distance information to model the similarity of the images and learning function maps from image pixels to a latent space, playing a vital role in rare object detection. In this paper, we propose a robust model based on a deep metric neural network for rare bioparticle (Cryptosporidium or Giardia) detection in drinking water. Experimental results showed that the deep metric neural network achieved a high accuracy of 99.86% in classification, 98.89% in precision rate, 99.16% in recall rate and zero false alarm rate. The reported model empowers imaging flow cytometry with capabilities of biomedical diagnosis, environmental monitoring, and other biosensing applications.

PhaseDCN: A Phase-Enhanced Dual-Path Dilated Convolutional Network for Single-Channel Speech Enhancement

Recent deep neural network (DNN) based single-channel speech enhancement methods have achieved remarkable results in the time-frequency (TF) magnitude domain. To further improve the quality and intelligibility of enhanced speech, the attention to phase enhancement is also increasing. In this paper, we propose a novel dilated convolutional network (DCN) model to simultaneously enhance the magnitude and phase of noisy speech. Unlike the direct complex spectral mapping methods, we take the complex spectrum of the signal as the main target and the ideal ratio mask (IRM) as the auxiliary target in a multi-target learning framework to achieve their complementary advantages. Firstly, a feature extraction module is introduced to achieve the fusion of local and long-term features. Two different targets are learned separately, but share the common feature extraction module, which is helpful to extract more general and suitable features. During the joint learning, the intermediate estimation of the IRM target in the auxiliary path, contributing as the attention gating factors, helps to distinguish the speech or non-speech components of the complex-valued signals in the main path. To leverage more fine-grained long-term contextual information, we introduce a multi-scale dilated convolution approach for feature encoding. Moreover, the proposed model is a causal system, which can fully meet the low latency requirements of real-time speech products. Experimental results show that, compared with other advanced systems, the proposed model not only has better speech denoising performance and phase estimation accuracy, but also generalizes better in the speaker, noise, and channel mismatch cases.

A proof that deep artificial neural networks overcome the curse of dimensionality in the numerical approximation of Kolmogorov partial differential equations with constant diffusion and nonlinear drift coefficients

In recent years deep artificial neural networks (DNNs) have been successfully employed in numerical simulations for a multitude of computational problems including, for example, object and face recognition, natural language processing, fraud detection, computational advertisement, and numerical approximations of partial differential equations (PDEs). These numerical simulations indicate that DNNs seem to possess the fundamental flexibility to overcome the curse of dimensionality in the sense that the number of real parameters used to describe the DNN grows at most polynomially in both the reciprocal of the prescribed approximation accuracy $ \varepsilon > 0 $ and the dimension $ d \in \mathbb{N}$ of the function which the DNN aims to approximate in such computational problems. There is also a large number of rigorous mathematical approximation results for artificial neural networks in the scientific literature but there are only a few special situations where results in the literature can rigorously justify the success of DNNs in high-dimensional function approximation. The key contribution of this paper is to reveal that DNNs do overcome the curse of dimensionality in the numerical approximation of Kolmogorov PDEs with constant diffusion and nonlinear drift coefficients. We prove that the number of parameters used to describe the employed DNN grows at most polynomially in both the PDE dimension $ d \in \mathbb{N}$ and the reciprocal of the prescribed approximation accuracy $ \varepsilon > 0 $. A crucial ingredient in our proof is the fact that the artificial neural network used to approximate the solution of the PDE is indeed a deep artificial neural network with a large number of hidden layers.

An Empirical Analysis of Deep Learning Architectures for Vehicle Make and Model Recognition

Vehicle make and model recognition is an important component of the Intelligent Transport System (ITS), which plays an essential role in the vehicle surveillance and traffic monitoring. In this paper, we evaluated the performance of the recent deep neural networks for vehicle recognition to identify 196 different types of vehicles based on their make, model, and year using Stanford Cars data. Transfer learning has been employed to reduce the training time and also added a dropout of 0.5 at the last dense layer before the output layer for improved generalization performance on an applied dataset. Different augmentation methods such as random erasing, horizontal flip, random crop, resize, rotate, shear, and sharpen were applied to the training dataset to increase the number of images. Besides, optimization methods namely mixup and cyclical learning rate were also investigated to train these deep neural networks to avoid overfitting. Finally, the proposed ensemble learning framework using k-fold cross-validation has resulted in the improved accuracy of image classification for the vehicle make and model recognition. The final achieved accuracy is 93.96% through ensembling homogeneous models of DenseNet201, which is higher than previous state-of-the-art approaches for the vehicle make and model recognition. Simultaneously, it has also improved the performance of other models where MobileNetV2 has shown the highest improvement with an increase of 2.8% in the overall accuracy using the proposed ensemble learning technique.

Reconstructing Spectra from RGB Images by Relative Error Least-Squares Regression

Spectral reconstruction (SR) algorithms attempt to map RGB- to hyperspectral-images. Classically, simple pixel-based regression is used to solve for this SR mapping and more recently patch-based Deep Neural Networks (DNN) are considered (with a modest performance increment). For either method, the 'training' process typically minimizes a Mean-Squared-Error (MSE) loss. Curiously, in recent research, SR algorithms are evaluated and ranked based on a relative percentage error, so-called MeanRelative-Absolute Error (MRAE), which behaves very differently from the MSE loss function. The most recent DNN approaches - perhaps unsurprisingly - directly optimize for this new MRAE error in training so as to match this new evaluation criteria.<br/> In this paper, we show how we can also reformulate pixelbased regression methods so that they too optimize a relative spectral error. Our Relative Error Least-Squares (RELS) approach minimizes an error that is similar to MRAE. Experiments demonstrate that regression models based on RELS deliver better spectral recovery, with up to a 10% increment in mean performance and a 20% improvement in worst-case performance depending on the method.