It has been consistently reported that many machine learning models are susceptible to adversarial attacks i.e., small additive adversarial perturbations applied to data points can cause misclassification. Adversarial training using empirical risk minimization is considered to be the state-of-the-art method for defense against adversarial attacks. Despite being successful in practice, several problems in understanding generalization performance of adversarial training remain open. In this paper, we derive precise theoretical predictions for the performance of adversarial training in binary classification. We consider the high-dimensional regime where the dimension of data grows with the size of the training data-set at a constant ratio. Our results provide exact asymptotics for standard and adversarial test errors of the estimators obtained by adversarial training with $\ell_q$-norm bounded perturbations ($q \ge 1$) for both discriminative binary models and generative Gaussian-mixture models with correlated features. Furthermore, we use these sharp predictions to uncover several intriguing observations on the role of various parameters including the over-parameterization ratio, the data model, and the attack budget on the adversarial and standard errors.
To date, there has been no formal study of the statistical cost of interpretability in machine learning. As such, the discourse around potential trade-offs is often informal and misconceptions abound. In this work, we aim to initiate a formal study of these trade-offs. A seemingly insurmountable roadblock is the lack of any agreed upon definition of interpretability. Instead, we propose a shift in perspective. Rather than attempt to define interpretability, we propose to model the \emph{act} of \emph{enforcing} interpretability. As a starting point, we focus on the setting of empirical risk minimization for binary classification, and view interpretability as a constraint placed on learning. That is, we assume we are given a subset of hypothesis that are deemed to be interpretable, possibly depending on the data distribution and other aspects of the context. We then model the act of enforcing interpretability as that of performing empirical risk minimization over the set of interpretable hypotheses. This model allows us to reason about the statistical implications of enforcing interpretability, using known results in statistical learning theory. Focusing on accuracy, we perform a case analysis, explaining why one may or may not observe a trade-off between accuracy and interpretability when the restriction to interpretable classifiers does or does not come at the cost of some excess statistical risk. We close with some worked examples and some open problems, which we hope will spur further theoretical development around the tradeoffs involved in interpretability.
Diabetic retinopathy (DR) is a retinal microvascular condition that emerges in diabetic patients. DR will continue to be a leading cause of blindness worldwide, with a predicted 191.0 million globally diagnosed patients in 2030. Microaneurysms, hemorrhages, exudates, and cotton wool spots are common signs of DR. However, they can be small and hard for human eyes to detect. Early detection of DR is crucial for effective clinical treatment. Existing methods to classify images require much time for feature extraction and selection, and are limited in their performance. Convolutional Neural Networks (CNNs), as an emerging deep learning (DL) method, have proven their potential in image classification tasks. In this paper, comprehensive experimental studies of implementing state-of-the-art CNNs for the detection and classification of DR are conducted in order to determine the top performing classifiers for the task. Five CNN classifiers, namely Inception-V3, VGG19, VGG16, ResNet50, and InceptionResNetV2, are evaluated through experiments. They categorize medical images into five different classes based on DR severity. Data augmentation and transfer learning techniques are applied since annotated medical images are limited and imbalanced. Experimental results indicate that the ResNet50 classifier has top performance for binary classification and that the InceptionResNetV2 classifier has top performance for multi-class DR classification.
Despite the wide empirical success of modern machine learning algorithms and models in a multitude of applications, they are known to be highly susceptible to seemingly small indiscernible perturbations to the input data known as \emph{adversarial attacks}. A variety of recent adversarial training procedures have been proposed to remedy this issue. Despite the success of such procedures at increasing accuracy on adversarially perturbed inputs or \emph{robust accuracy}, these techniques often reduce accuracy on natural unperturbed inputs or \emph{standard accuracy}. Complicating matters further, the effect and trend of adversarial training procedures on standard and robust accuracy is rather counter intuitive and radically dependent on a variety of factors including the perceived form of the perturbation during training, size/quality of data, model overparameterization, etc. In this paper we focus on binary classification problems where the data is generated according to the mixture of two Gaussians with general anisotropic covariance matrices and derive a precise characterization of the standard and robust accuracy for a class of minimax adversarially trained models. We consider a general norm-based adversarial model, where the adversary can add perturbations of bounded $\ell_p$ norm to each input data, for an arbitrary $p\ge 1$. Our comprehensive analysis allows us to theoretically explain several intriguing empirical phenomena and provide a precise understanding of the role of different problem parameters on standard and robust accuracies.
The traditional binary classification framework constructs classifiers which may have good accuracy, but whose false positive and false negative error rates are not under users' control. In many cases, one of the errors is more severe and only the classifiers with the corresponding rate lower than the predefined threshold are acceptable. In this study, we combine binary classification with statistical hypothesis testing to control the target error rate of already trained classifiers. In particular, we show how to turn binary classifiers into statistical tests, calculate the classification p-values, and use them to limit the target error rate.
Automatic visual fire detection is used to complement traditional fire detection sensor systems (smoke/heat). In this work, we investigate different Convolutional Neural Network (CNN) architectures and their variants for the non-temporal real-time bounds detection of fire pixel regions in video (or still) imagery. Two reduced complexity compact CNN architectures (NasNet-A-OnFire and ShuffleNetV2-OnFire) are proposed through experimental analysis to optimise the computational efficiency for this task. The results improve upon the current state-of-the-art solution for fire detection, achieving an accuracy of 95% for full-frame binary classification and 97% for superpixel localisation. We notably achieve a classification speed up by a factor of 2.3x for binary classification and 1.3x for superpixel localisation, with runtime of 40 fps and 18 fps respectively, outperforming prior work in the field presenting an efficient, robust and real-time solution for fire region detection. Subsequent implementation on low-powered devices (Nvidia Xavier-NX, achieving 49 fps for full-frame classification via ShuffleNetV2-OnFire) demonstrates our architectures are suitable for various real-world deployment applications.
We use Generative Adversarial Networks (GANs) to design a class conditional label noise (CCN) robust scheme for binary classification. It first generates a set of correctly labelled data points from noisy labelled data and 0.1% or 1% clean labels such that the generated and true (clean) labelled data distributions are close; generated labelled data is used to learn a good classifier. The mode collapse problem while generating correct feature-label pairs and the problem of skewed feature-label dimension ratio ($\sim$ 784:1) are avoided by using Wasserstein GAN and a simple data representation change. Another WGAN with information-theoretic flavour on top of the new representation is also proposed. The major advantage of both schemes is their significant improvement over the existing ones in presence of very high CCN rates, without either estimating or cross-validating over the noise rates. We proved that KL divergence between clean and noisy distribution increases w.r.t. noise rates in symmetric label noise model; can be extended to high CCN rates. This implies that our schemes perform well due to the adversarial nature of GANs. Further, use of generative approach (learning clean joint distribution) while handling noise enables our schemes to perform better than discriminative approaches like GLC, LDMI and GCE; even when the classes are highly imbalanced. Using Friedman F test and Nemenyi posthoc test, we showed that on high dimensional binary class synthetic, MNIST and Fashion MNIST datasets, our schemes outperform the existing methods and demonstrate consistent performance across noise rates.
We present a greedy algorithm for solving binary classification problems in situations where the dataset is either too small or not fully representative of the problem being solved, and obtaining more data is not possible. This algorithm is of particular interest when training small models that have trouble generalizing. It relies on a trained model with loose accuracy constraints, an iterative hyperparameter pruning procedure, and a function used to generate new data. Analysis on correctness and runtime complexity under ideal conditions and an extension to deep neural networks is provided. In the former case we obtain an asymptotic bound of $O\left(|\Theta^2|\left(\log{|\Theta|} + |\theta^2| + T_f\left(| D|\right)\right) + \bar{S}|\Theta||{E}|\right)$, where $|{\Theta}|$ is the cardinality of the set of hyperparameters $\theta$ to be searched; $|{E}|$ and $|{D}|$ are the sizes of the evaluation and training datasets, respectively; $\bar{S}$ and $\bar{f}$ are the inference times for the trained model and the candidate model; and $T_f({|{D}|})$ is a polynomial on $|{D}|$ and $\bar{f}$. Under these conditions, this algorithm returns a solution that is $1 \leq r \leq 2(1 - {2^{-|{\Theta}|}})$ times better than simply enumerating and training with any $\theta \in \Theta$. As part of our analysis of the generating function we also prove that, under certain assumptions, if an open cover of $D$ has the same homology as the manifold where the support of the underlying probability distribution lies, then $D$ is learnable, and viceversa.
Cyberbullying is a prevalent social problem that inflicts detrimental consequences to the health and safety of victims such as psychological distress, anti-social behaviour, and suicide. The automation of cyberbullying detection is a recent but widely researched problem, with current research having a strong focus on a binary classification of bullying versus non-bullying. This paper proposes a novel approach to enhancing cyberbullying detection through role modeling. We utilise a dataset from ASKfm to perform multi-class classification to detect participant roles (e.g. victim, harasser). Our preliminary results demonstrate promising performance including 0.83 and 0.76 of F1-score for cyberbullying and role classification respectively, outperforming baselines.
The task of aesthetic quality assessment is complicated due to its subjectivity. In recent years, the target representation of image aesthetic quality has changed from a one-dimensional binary classification label or numerical score to a multi-dimensional score distribution. According to current methods, the ground truth score distributions are straightforwardly regressed. However, the subjectivity of aesthetics is not taken into account, that is to say, the psychological processes of human beings are not taken into consideration, which limits the performance of the task. In this paper, we propose a Deep Drift-Diffusion (DDD) model inspired by psychologists to predict aesthetic score distribution from images. The DDD model can describe the psychological process of aesthetic perception instead of traditional modeling of the results of assessment. We use deep convolution neural networks to regress the parameters of the drift-diffusion model. The experimental results in large scale aesthetic image datasets reveal that our novel DDD model is simple but efficient, which outperforms the state-of-the-art methods in aesthetic score distribution prediction. Besides, different psychological processes can also be predicted by our model.
The ability to correctly model distinct meanings of a word is crucial for the effectiveness of semantic representation techniques. However, most existing evaluation benchmarks for assessing this criterion are tied to sense inventories (usually WordNet), restricting their usage to a small subset of knowledge-based representation techniques. The Word-in-Context dataset (WiC) addresses the dependence on sense inventories by reformulating the standard disambiguation task as a binary classification problem; but, it is limited to the English language. We put forward a large multilingual benchmark, XL-WiC, featuring gold standards in 12 new languages from varied language families and with different degrees of resource availability, opening room for evaluation scenarios such as zero-shot cross-lingual transfer. We perform a series of experiments to determine the reliability of the datasets and to set performance baselines for several recent contextualized multilingual models. Experimental results show that even when no tagged instances are available for a target language, models trained solely on the English data can attain competitive performance in the task of distinguishing different meanings of a word, even for distant languages. XL-WiC is available at https://pilehvar.github.io/xlwic/.
One reason pretraining on self-supervised linguistic tasks is effective is that it teaches models features that are helpful for language understanding. However, we want pretrained models to learn not only to represent linguistic features, but also to use those features preferentially during fine-turning. With this goal in mind, we introduce a new English-language diagnostic set called MSGS (the Mixed Signals Generalization Set), which consists of 20 ambiguous binary classification tasks that we use to test whether a pretrained model prefers linguistic or surface generalizations during fine-tuning. We pretrain RoBERTa models from scratch on quantities of data ranging from 1M to 1B words and compare their performance on MSGS to the publicly available RoBERTa-base. We find that models can learn to represent linguistic features with little pretraining data, but require far more data to learn to prefer linguistic generalizations over surface ones. Eventually, with about 30B words of pretraining data, RoBERTa-base does demonstrate a linguistic bias with some regularity. We conclude that while self-supervised pretraining is an effective way to learn helpful inductive biases, there is likely room to improve the rate at which models learn which features matter.
Large pre-trained language representation models (LMs) have recently collected a huge number of successes in many NLP tasks. In 2018 BERT, and later its successors (e.g. RoBERTa), obtained state-of-the-art results in classical benchmark tasks, such as GLUE benchmark. After that, works about adversarial attacks have been published to test their generalization proprieties and robustness. In this work, we design Evolutionary Fooling Sentences Generator (EFSG), a model- and task-agnostic adversarial attack algorithm built using an evolutionary approach to generate false-positive sentences for binary classification tasks. We successfully apply EFSG to CoLA and MRPC tasks, on BERT and RoBERTa, comparing performances. Results prove the presence of weak spots in state-of-the-art LMs. We finally test adversarial training as a data augmentation defence approach against EFSG, obtaining stronger improved models with no loss of accuracy when tested on the original datasets.
For several years till date, the major issues in terms of solving for classification problems are the issues of Imbalanced data. Because majority of the machine learning algorithms by default assumes all data are balanced, the algorithms do not take into consideration the distribution of the data sample class. The results tend to be unsatisfactory and skewed towards the majority sample class distribution. This implies that the consequences as a result of using a model built using an Imbalanced data without handling for the Imbalance in the data could be misleading both in practice and theory. Most researchers have focused on the application of Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) Sampling Approach in handling data Imbalance independently in their works and have failed to better explain the algorithms behind these techniques with computed examples. This paper focuses on both synthetic oversampling techniques and manually computes synthetic data points to enhance easy comprehension of the algorithms. We analyze the application of these synthetic oversampling techniques on binary classification problems with different Imbalanced ratios and sample sizes.
Systematic reviews, which entail the extraction of data from large numbers of scientific documents, are an ideal avenue for the application of machine learning. They are vital to many fields of science and philanthropy, but are very time-consuming and require experts. Yet the three main stages of a systematic review are easily done automatically: searching for documents can be done via APIs and scrapers, selection of relevant documents can be done via binary classification, and extraction of data can be done via sequence-labelling classification. Despite the promise of automation for this field, little research exists that examines the various ways to automate each of these tasks. We construct a pipeline that automates each of these aspects, and experiment with many human-time vs. system quality trade-offs. We test the ability of classifiers to work well on small amounts of data and to generalise to data from countries not represented in the training data. We test different types of data extraction with varying difficulty in annotation, and five different neural architectures to do the extraction. We find that we can get surprising accuracy and generalisability of the whole pipeline system with only 2 weeks of human-expert annotation, which is only 15% of the time it takes to do the whole review manually and can be repeated and extended to new data with no additional effort.
In this paper, we propose a novel approach for implicit data representation to evaluate similarity of input data using a trained neural network. In contrast to the previous approach, which uses gradients for representation, we utilize only the outputs from the last hidden layer of a neural network and do not use a backward step. The proposed technique explicitly takes into account the initial task and significantly reduces the size of the vector representation, as well as the computation time. The key point is minimization of information loss between layers. Generally, a neural network discards information that is not related to the problem, which makes the last hidden layer representation useless for input similarity task. In this work, we consider two main causes of information loss: correlation between neurons and insufficient size of the last hidden layer. To reduce the correlation between neurons we use orthogonal weight initialization for each layer and modify the loss function to ensure orthogonality of the weights during training. Moreover, we show that activation functions can potentially increase correlation. To solve this problem, we apply modified Batch-Normalization with Dropout. Using orthogonal weight matrices allow us to consider such neural networks as an application of the Random Projection method and get a lower bound estimate for the size of the last hidden layer. We perform experiments on MNIST and physical examination datasets. In both experiments, initially, we split a set of labels into two disjoint subsets to train a neural network for binary classification problem, and then use this model to measure similarity between input data and define hidden classes. Our experimental results show that the proposed approach achieves competitive results on the input similarity task while reducing both computation time and the size of the input representation.
As Machine Learning technologies become increasingly used in contexts that affect citizens, companies as well as researchers need to be confident that their application of these methods will not have unexpected social implications, such as bias towards gender, ethnicity, and/or people with disabilities. There is significant literature on approaches to mitigate bias and promote fairness, yet the area is complex and hard to penetrate for newcomers to the domain. This article seeks to provide an overview of the different schools of thought and approaches to mitigating (social) biases and increase fairness in the Machine Learning literature. It organises approaches into the widely accepted framework of pre-processing, in-processing, and post-processing methods, subcategorizing into a further 11 method areas. Although much of the literature emphasizes binary classification, a discussion of fairness in regression, recommender systems, unsupervised learning, and natural language processing is also provided along with a selection of currently available open source libraries. The article concludes by summarising open challenges articulated as four dilemmas for fairness research.
Kernel approximation is widely used to scale up kernel SVM training and prediction. However, the memory and computation costs of kernel approximation models are still too high if we want to deploy them on memory-limited devices such as mobile phones, smartwatches, and IoT devices. To address this challenge, we propose a novel memory and computation-efficient kernel SVM model by using both binary embedding and binary model coefficients. First, we propose an efficient way to generate compact binary embedding of the data, preserving the kernel similarity. Second, we propose a simple but effective algorithm to learn a linear classification model with ternary coefficients that can support different types of loss function and regularizer. Our algorithm can achieve better generalization accuracy than existing works on learning binary coefficients since we allow coefficient to be $-1$, $0$, or $1$ during the training stage, and coefficient $0$ can be removed during model inference for binary classification. Moreover, we provide a detailed analysis of the convergence of our algorithm and the inference complexity of our model. The analysis shows that the convergence to a local optimum is guaranteed, and the inference complexity of our model is much lower than other competing methods. Our experimental results on five large real-world datasets have demonstrated that our proposed method can build accurate nonlinear SVM models with memory costs less than 30KB.
In forming learning objectives, one oftentimes needs to aggregate a set of individual values to a single output. Such cases occur in the aggregate loss, which combines individual losses of a learning model over each training sample, and in the individual loss for multi-label learning, which combines prediction scores over all class labels. In this work, we introduce the sum of ranked range (SoRR) as a general approach to form learning objectives. A ranked range is a consecutive sequence of sorted values of a set of real numbers. The minimization of SoRR is solved with the difference of convex algorithm (DCA). We explore two applications in machine learning of the minimization of the SoRR framework, namely the AoRR aggregate loss for binary classification and the TKML individual loss for multi-label/multi-class classification. Our empirical results highlight the effectiveness of the proposed optimization framework and demonstrate the applicability of proposed losses using synthetic and real datasets.
To alleviate the data requirement for training effective binary classifiers in binary classification, many weakly supervised learning settings have been proposed. Among them, some consider using pairwise but not pointwise labels, when pointwise labels are not accessible due to privacy, confidentiality, or security reasons. However, as a pairwise label denotes whether or not two data points share a pointwise label, it cannot be easily collected if either point is equally likely to be positive or negative. Thus, in this paper, we propose a novel setting called pairwise comparison (Pcomp) classification, where we have only pairs of unlabeled data that we know one is more likely to be positive than the other. Firstly, we give a Pcomp data generation process, derive an unbiased risk estimator (URE) with theoretical guarantee, and further improve URE using correction functions. Secondly, we link Pcomp classification to noisy-label learning to develop a progressive URE and improve it by imposing consistency regularization. Finally, we demonstrate by experiments the effectiveness of our methods, which suggests Pcomp is a valuable and practically useful type of pairwise supervision besides the pairwise label.