In this report we propose a classification technique for skin lesion images as a part of our submission for ISIC 2018 Challenge in Skin Lesion Analysis Towards Melanoma Detection. Our data was extracted from the ISIC 2018: Skin Lesion Analysis Towards Melanoma Detection grand challenge datasets. The features are extracted through a Convolutional Neural Network, in our case ResNet50 and then using these features we train a DeepForest, having cascading layers, to classify our skin lesion images. We know that Convolutional Neural Networks are a state-of-the-art technique in representation learning for images, with the convolutional filters learning to detect features from images through backpropagation. These features are then usually fed to a classifier like a softmax layer or other such classifiers for classification tasks. In our case we do not use the traditional backpropagation method and train a softmax layer for classification. Instead, we use Deep Forest, a novel decision tree ensemble approach with performance highly competitive to deep neural networks in a broad range of tasks. Thus we use a ResNet50 to extract the features from skin lesion images and then use the Deep Forest to classify these images. This method has been used because Deep Forest has been found to be hugely efficient in areas where there are only small-scale training data available. Also as the Deep Forest network decides its complexity by itself, it also caters to the problem of dataset imbalance we faced in this problem.
In this work, we focus on effectively leveraging and integrating information from concept-level as well as word-level via projecting concepts and words into a lower dimensional space while retaining most critical semantics. In a broad context of opinion understanding system, we investigate the use of the fused embedding for several core NLP tasks: named entity detection and classification, automatic speech recognition reranking, and targeted sentiment analysis.
This work summarizes our submission for the Task 3: Disease Classification of ISIC 2018 challenge in Skin Lesion Analysis Towards Melanoma Detection. We use a novel deep neural network (DNN) ensemble architecture introduced by us that can effectively classify skin lesions by using data-augmentation and bagging to address paucity of data and prevent over-fitting. The ensemble is composed of two DNN architectures: Inception-v4 and Inception-Resnet-v2. The DNN architectures are combined in to an ensemble by using a $1\times1$ convolution for fusion in a meta-learning layer.
In this paper, an image denoising algorithm is proposed for salt and pepper noise. First, a generative model is built on a patch as a basic unit and then the algorithm locates the image noise within that patch in order to better describe the patch and obtain better subsequent clustering. Second, the algorithm classifies patches using a generative clustering method, thus providing additional similarity information for noise repair and suppressing the interference of noise, abandoning those categories that consist of a smaller number of patches. Finally, the algorithm builds a non-local switching filter to remove the salt and pepper noise. Simulation results show that the proposed algorithm effectively denoises salt and pepper noise of various densities. It obtains a better visual quality and higher peak signal-to-noise ratio score than several state-of-the-art algorithms. In short, our algorithm uses a noisy patch as the basic unit, a patch clustering method to optimize the repair data set as well as obtain a better denoising effect, and provides a guideline for future denoising and repair methods.
This paper advances the state of the art in human examination of iris images by (1) assessing the impact of different iris conditions in identity verification, and (2) introducing an annotation step that improves the accuracy of people's decisions. In a first experimental session, 114 subjects were asked to decide if pairs of iris images depict the same eye (genuine pairs) or two distinct eyes (impostor pairs). The image pairs sampled six conditions: (1) easy for algorithms to classify, (2) difficult for algorithms to classify, (3) large difference in pupil dilation, (4) disease-affected eyes, (5) identical twins, and (6) post-mortem samples. In a second session, 85 of the 114 subjects were asked to annotate matching and non-matching regions that supported their decisions. Subjects were allowed to change their initial classification as a result of the annotation process. Results suggest that: (a) people improve their identity verification accuracy when asked to annotate matching and non-matching regions between the pair of images, (b) images depicting the same eye with large difference in pupil dilation were the most challenging to subjects, but benefited well from the annotation-driven classification, (c) humans performed better than iris recognition algorithms when verifying genuine pairs of post-mortem and disease-affected eyes (i.e., samples showing deformations that go beyond the distortions of a healthy iris due to pupil dilation), and (d) annotation does not improve accuracy of analyzing images from identical twins, which remain confusing for people.
Machine learning models leak information about the datasets on which they are trained. An adversary can build an algorithm to trace the individual members of a model's training dataset. As a fundamental inference attack, he aims to distinguish between data points that were part of the model's training set and any other data points from the same distribution. This is known as the tracing (and also membership inference) attack. In this paper, we focus on such attacks against black-box models, where the adversary can only observe the output of the model, but not its parameters. This is the current setting of machine learning as a service in the Internet. We introduce a privacy mechanism to train machine learning models that provably achieve membership privacy: the model's predictions on its training data are indistinguishable from its predictions on other data points from the same distribution. We design a strategic mechanism where the privacy mechanism anticipates the membership inference attacks. The objective is to train a model such that not only does it have the minimum prediction error (high utility), but also it is the most robust model against its corresponding strongest inference attack (high privacy). We formalize this as a min-max game optimization problem, and design an adversarial training algorithm that minimizes the classification loss of the model as well as the maximum gain of the membership inference attack against it. This strategy, which guarantees membership privacy (as prediction indistinguishability), acts also as a strong regularizer and significantly generalizes the model. We evaluate our privacy mechanism on deep neural networks using different benchmark datasets. We show that our min-max strategy can mitigate the risk of membership inference attacks (close to the random guess) with a negligible cost in terms of the classification error.
Image classification is an ongoing research challenge. Most of the available research focuses on image classification for the English language, however there is very little research on image classification for the Arabic language. Expanding image classification to Arabic has several applications. The present study investigated a method for generating Arabic labels for images of objects. The method used in this study involved a direct English to Arabic translation of the labels that are currently available on ImageNet, a database commonly used in image classification research. The purpose of this study was to test the accuracy of this method. In this study, 2,887 labeled images were randomly selected from ImageNet. All of the labels were translated from English to Arabic using Google Translate. The accuracy of the translations was evaluated. Results indicated that that 65.6% of the Arabic labels were accurate. This study makes three important contributions to the image classification literature: (1) it determined the baseline level of accuracy for algorithms that provide Arabic labels for images, (2) it provided 1,895 images that are tagged with accurate Arabic labels, and (3) provided the accuracy of translations of image labels from English to Arabic.
Deep learning techniques have recently shown to be successful in many natural language processing tasks forming state-of-the-art systems. They require, however, a large amount of annotated data which is often missing. This paper explores the use of domain-adversarial learning as a regularizer to avoid overfitting when training domain invariant features for deep, complex neural networks in low-resource and zero-resource settings in new target domains or languages. In case of new languages, we show that monolingual word vectors can be directly used for training without prealignment. Their projection into a common space can be learnt ad-hoc at training time reaching the final performance of pretrained multilingual word vectors.
Feature attribution methods, or saliency maps, are one of the most popular approaches for explaining the decisions of complex machine learning models such as deep neural networks. In this study, we propose a stochastic optimization approach for the perturbation-based feature attribution method. While the original optimization problem of the perturbation-based feature attribution is difficult to solve because of the complex constraints, we propose to reformulate the problem as the maximization of a differentiable function, which can be solved using gradient-based algorithms. In particular, stochastic optimization is well-suited for the proposed reformulation, and we can solve the problem using popular algorithms such as SGD, RMSProp, and Adam. The experiment on the image classification with VGG16 shows that the proposed method could identify relevant parts of the images effectively.
Deep learning approaches for sentiment classification do not fully exploit sentiment linguistic knowledge. In this paper, we propose a Multi-sentiment-resource Enhanced Attention Network (MEAN) to alleviate the problem by integrating three kinds of sentiment linguistic knowledge (e.g., sentiment lexicon, negation words, intensity words) into the deep neural network via attention mechanisms. By using various types of sentiment resources, MEAN utilizes sentiment-relevant information from different representation subspaces, which makes it more effective to capture the overall semantics of the sentiment, negation and intensity words for sentiment prediction. The experimental results demonstrate that MEAN has robust superiority over strong competitors.
It is desirable for detection and classification algorithms to generalize to unfamiliar environments, but suitable benchmarks for quantitatively studying this phenomenon are not yet available. We present a dataset designed to measure recognition generalization to novel environments. The images in our dataset are harvested from twenty camera traps deployed to monitor animal populations. Camera traps are fixed at one location, hence the background changes little across images; capture is triggered automatically, hence there is no human bias. The challenge is learning recognition in a handful of locations, and generalizing animal detection and classification to new locations where no training data is available. In our experiments state-of-the-art algorithms show excellent performance when tested at the same location where they were trained. However, we find that generalization to new locations is poor, especially for classification systems.
A discriminative structured analysis dictionary is proposed for the classification task. A structure of the union of subspaces (UoS) is integrated into the conventional analysis dictionary learning to enhance the capability of discrimination. A simple classifier is also simultaneously included into the formulated functional to ensure a more complete consistent classification. The solution of the algorithm is efficiently obtained by the linearized alternating direction method of multipliers. Moreover, a distributed structured analysis dictionary learning is also presented to address large scale datasets. It can group-(class-) independently train the structured analysis dictionaries by different machines/cores/threads, and therefore avoid a high computational cost. A consensus structured analysis dictionary and a global classifier are jointly learned in the distributed approach to safeguard the discriminative power and the efficiency of classification. Experiments demonstrate that our method achieves a comparable or better performance than the state-of-the-art algorithms in a variety of visual classification tasks. In addition, the training and testing computational complexity are also greatly reduced.
In this article we present a novel underwater dataset collected from several field trials within the EU FP7 project "Cognitive autonomous diving buddy (CADDY)", where an Autonomous Underwater Vehicle (AUV) was used to interact with divers and monitor their activities. To our knowledge, this is one of the first efforts to collect a large dataset in underwater environments targeting object classification, segmentation and human pose estimation tasks. The first part of the dataset contains stereo camera recordings (~10K) of divers performing hand gestures to communicate and interact with an AUV in different environmental conditions. These gestures samples serve to test the robustness of object detection and classification algorithms against underwater image distortions i.e., color attenuation and light backscatter. The second part includes stereo footage (~12.7K) of divers free-swimming in front of the AUV, along with synchronized IMUs measurements located throughout the diver's suit (DiverNet) which serve as ground-truth for human pose and tracking methods. In both cases, these rectified images allow investigation of 3D representation and reasoning pipelines from low-texture targets commonly present in underwater scenarios. In this paper we describe our recording platform, sensor calibration procedure plus the data format and the utilities provided to use the dataset.
As known, attribute selection is a method that is used before the classification of data mining. In this study, a new data set has been created by using attributes expressing overall satisfaction in Turkey Statistical Institute (TSI) Life Satisfaction Survey dataset. Attributes are sorted by Ranking search method using attribute selection algorithms in a data mining application. These selected attributes were subjected to a classification test with Naive Bayes and Random Forest from machine learning algorithms. The feature selection algorithms are compared according to the number of attributes selected and the classification accuracy rates achievable with them. In this study, which is aimed at reducing the dataset volume, the best classification result comes up with 3 attributes selected by the Chi2 algorithm. The best classification rate was 73% with the Random Forest classification algorithm.
Various and ubiquitous information systems are being used in monitoring, exchanging, and collecting information. These systems are generating massive amount of event sequence logs that may help us understand underlying phenomenon. By analyzing these logs, we can learn process models that describe system procedures, predict the development of the system, or check whether the changes are expected. In this paper, we consider a novel technique that models these sequences of events in temporal-probabilistic manners. Specifically, we propose a probabilistic process model that combines hidden semi-Markov model and classification trees learning. Our experimental result shows that the proposed approach can answer a kind of question-"what are the most frequent sequence of system dynamics relevant to a given sequence of observable events?". For example, "Given a series of medical treatments, what are the most relevant patients' health condition pattern changes at different times?"
In text classification, the problem of overfitting arises due to the high dimensionality, making regularization essential. Although classic regularizers provide sparsity, they fail to return highly accurate models. On the contrary, state-of-the-art group-lasso regularizers provide better results at the expense of low sparsity. In this paper, we apply a greedy variable selection algorithm, called Orthogonal Matching Pursuit, for the text classification task. We also extend standard group OMP by introducing overlapping Group OMP to handle overlapping groups of features. Empirical analysis verifies that both OMP and overlapping GOMP constitute powerful regularizers, able to produce effective and very sparse models. Code and data are available online: https://github.com/y3nk0/OMP-for-Text-Classification .
Recently proposed techniques for semi-supervised learning such as Temporal Ensembling and Mean Teacher have achieved state-of-the-art results in many important classification benchmarks. In this work, we expand the Mean Teacher approach to segmentation tasks and show that it can bring important improvements in a realistic small data regime using a publicly available multi-center dataset from the Magnetic Resonance Imaging (MRI) domain. We also devise a method to solve the problems that arise when using traditional data augmentation strategies for segmentation tasks on our new training scheme.
Without any specific way for imbalance data classification, artificial intelligence algorithm cannot recognize data from minority classes easily. In general, modifying the existing algorithm by assuming that the training data is imbalanced, is the only way to handle imbalance data. However, for a normal data handling, this way mostly produces a deficient result. In this research, we propose a class expert generative adversarial network (CE-GAN) as the solution for imbalance data classification. CE-GAN is a modification in deep learning algorithm architecture that does not have an assumption that the training data is imbalance data. Moreover, CE-GAN is designed to identify more detail about the character of each class before classification step. CE-GAN has been proved in this research to give a good performance for imbalance data classification.
Post-mortem iris recognition can offer an additional forensic method of personal identification. However, in contrary to already well-established human examination of fingerprints, making iris recognition human-interpretable is harder, and therefore it has never been applied in forensic proceedings. There is no strong consensus among biometric experts which iris features, especially those in iris images acquired post-mortem, are the most important for human experts solving an iris recognition task. This paper explores two ways of broadening this knowledge: (a) with an eye tracker, the salient features used by humans comparing iris images on a screen are extracted, and (b) class-activation maps produced by the convolutional neural network solving the iris recognition task are analyzed. Both humans and deep learning-based solutions were examined with the same set of iris image pairs. This made it possible to compare the attention maps and conclude that (a) deep learning-based method can offer human-interpretable decisions backed by visual explanations pointing a human examiner to salient regions, and (b) in many cases humans and a machine used different features, what means that a deep learning-based method can offer a complementary support to human experts. This paper offers the first known to us human-interpretable comparison of machine-based and human-based post-mortem iris recognition, and the trained models annotating salient iris image regions.
In this paper, we present our deep attention-based classification (DABC) network for robust single image depth prediction, in the context of the Robust Vision Challenge 2018 (ROB 2018). Unlike conventional depth prediction, our goal is to design a model that can perform well in both indoor and outdoor scenes with a single parameter set. However, robust depth prediction suffers from two challenging problems: a) How to extract more discriminative features for different scenes (compared to a single scene)? b) How to handle the large differences of depth ranges between indoor and outdoor datasets? To address these two problems, we first formulate depth prediction as a multi-class classification task and apply a softmax classifier to classify the depth label of each pixel. We then introduce a global pooling layer and a channel-wise attention mechanism to adaptively select the discriminative channels of features and to update the original features by assigning important channels with higher weights. Further, to reduce the influence of quantization errors, we employ a soft-weighted sum inference strategy for the final prediction. Experimental results on both indoor and outdoor datasets demonstrate the effectiveness of our method. It is worth mentioning that we won the 2-nd place in single image depth prediction entry of ROB 2018, in conjunction with IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018.