We focus our attention on the link prediction problem for knowledge graphs, which is treated herein as a binary classification task on neural embeddings of the entities. By comparing, combining and extending different methodologies for link prediction on graph-based data coming from different domains, we formalize a unified methodology for the quality evaluation benchmark of neural embeddings for knowledge graphs. This benchmark is then used to empirically investigate the potential of training neural embeddings globally for the entire graph, as opposed to the usual way of training embeddings locally for a specific relation. This new way of testing the quality of the embeddings evaluates the performance of binary classifiers for scalable link prediction with limited data. Our evaluation pipeline is made open source, and with this we aim to draw more attention of the community towards an important issue of transparency and reproducibility of the neural embeddings evaluations.
We present the problem of selecting relevant premises for a proof of a given statement. When stated as a binary classification task for pairs (conjecture, axiom), it can be efficiently solved using artificial neural networks. The key difference between our advance to solve this problem and previous approaches is the use of just functional signatures of premises. To further improve the performance of the model, we use dimensionality reduction technique, to replace long and sparse signature vectors with their compact and dense embedded versions. These are obtained by firstly defining the concept of a context for each functor symbol, and then training a simple neural network to predict the distribution of other functor symbols in the context of this functor. After training the network, the output of its hidden layer is used to construct a lower dimensional embedding of a functional signature (for each premise) with a distributed representation of features. This allows us to use 512-dimensional embeddings for conjecture-axiom pairs, containing enough information about the original statements to reach the accuracy of 76.45% in premise selection task, only with simple two-layer densely connected neural networks.
Adversarial perturbations are noise-like patterns that can subtly change the data, while failing an otherwise accurate classifier. In this paper, we propose to use such perturbations within a novel contrastive learning setup to build negative samples, which are then used to produce improved video representations. To this end, given a well-trained deep model for per-frame video recognition, we first generate adversarial noise adapted to this model. Positive and negative bags are produced using the original data features from the full video sequence and their perturbed counterparts, respectively. Unlike the classic contrastive learning methods, we develop a binary classification problem that learns a set of discriminative hyperplanes -- as a subspace -- that will separate the two bags from each other. This subspace is then used as a descriptor for the video, dubbed \emph{discriminative subspace pooling}. As the perturbed features belong to data classes that are likely to be confused with the original features, the discriminative subspace will characterize parts of the feature space that are more representative of the original data, and thus may provide robust video representations. To learn such descriptors, we formulate a subspace learning objective on the Stiefel manifold and resort to Riemannian optimization methods for solving it efficiently. We provide experiments on several video datasets and demonstrate state-of-the-art results.
There is a recent interest in developing statistical filtering methods for stochastic optimization (FSO) by leveraging a probabilistic perspective of incremental proximity methods (IPMs). The existent FSO methods are derived based on the Kalman filter (KF) and extended KF (EKF). Different from classical stochastic optimization methods such as the stochastic gradient descent (SGD) and typical IPMs, such KF-type algorithms possess a desirable property, namely they do not require pre-scheduling of the learning rate for convergence. However, on the other side, they have inherent limitations inherited from the nature of KF mechanisms. It is a consensus that the class of particle filters (PFs) outperforms the KF and its variants remarkably for nonlinear and/or non-Gaussian statistical filtering tasks. Hence, it is natural to ask if the FSO methods can benefit from the PF theory to get around of limitations of the KF-type stochastic optimization methods. We provide an affirmative answer to the aforementioned question by developing two PF based stochastic optimizers (PFSOs). For performance evaluation, we apply them to address nonlinear least-square fitting using simulated data sets and empirical risk minimization for binary classification using real data sets. Experimental results demonstrate that PFSOs outperform remarkably existent methods in terms of numerical stability, convergence speed, and flexibility in handling different types of loss functions.
Many problems that appear in biomedical decision making, such as diagnosing disease and predicting response to treatment, can be expressed as binary classification problems. The costs of false positives and false negatives vary across application domains and receiver operating characteristic (ROC) curves provide a visual representation of this trade-off. Nonparametric estimators for the ROC curve, such as a weighted support vector machine (SVM), are desirable because they are robust to model misspecification. While weighted SVMs have great potential for estimating ROC curves, their theoretical properties were heretofore underdeveloped. We propose a method for constructing confidence bands for the SVM ROC curve and provide the theoretical justification for the SVM ROC curve by showing that the risk function of the estimated decision rule is uniformly consistent across the weight parameter. We demonstrate the proposed confidence band method and the superior sensitivity and specificity of the weighted SVM compared to commonly used methods in diagnostic medicine using simulation studies. We present two illustrative examples: diagnosis of hepatitis C and a predictive model for treatment response in breast cancer.
We examine the supervised learning problem in its continuous setting and give a general optimality condition through techniques of functional analysis and the calculus of variations. This enables us to solve the optimality condition for the desired function u numerically and make several comparisons with other widely utilized supervised learning models. We employ the accuracy and area under the receiver operating characteristic curve as metrics of the performance. Finally, 3 analyses are conducted based on these two mentioned metrics where we compare the models and make conclusions to determine whether or not our method is competitive.
Ventricular Fibrillation (VF), one of the most dangerous arrhythmias, is responsible for sudden cardiac arrests. Thus, various algorithms have been developed to predict VF from Electrocardiogram (ECG), which is a binary classification problem. In the literature, we find a number of algorithms based on signal processing, where, after some robust mathematical operations the decision is given based on a predefined threshold over a single value. On the other hand, some machine learning based algorithms are also reported in the literature; however, these algorithms merely combine some parameters and make a prediction using those as features. Both the approaches have their perks and pitfalls; thus our motivation was to coalesce them to get the best out of the both worlds. Hence we have developed, VFPred that, in addition to employing a signal processing pipeline, namely, Empirical Mode Decomposition and Discrete Time Fourier Transform for useful feature extraction, uses a Support Vector Machine for efficient classification. VFPred turns out to be a robust algorithm as it is able to successfully segregate the two classes with equal confidence (Sensitivity = 99.99%, Specificity = 98.40%) even from a short signal of 5 seconds long, whereas existing works though requires longer signals, flourishes in one but fails in the other.
The electrocardiogram or ECG has been in use for over 100 years and remains the most widely performed diagnostic test to characterize cardiac structure and electrical activity. We hypothesized that parallel advances in computing power, innovations in machine learning algorithms, and availability of large-scale digitized ECG data would enable extending the utility of the ECG beyond its current limitations, while at the same time preserving interpretability, which is fundamental to medical decision-making. We identified 36,186 ECGs from the UCSF database that were 1) in normal sinus rhythm and 2) would enable training of specific models for estimation of cardiac structure or function or detection of disease. We derived a novel model for ECG segmentation using convolutional neural networks (CNN) and Hidden Markov Models (HMM) and evaluated its output by comparing electrical interval estimates to 141,864 measurements from the clinical workflow. We built a 725-element patient-level ECG profile using downsampled segmentation data and trained machine learning models to estimate left ventricular mass, left atrial volume, mitral annulus e' and to detect and track four diseases: pulmonary arterial hypertension (PAH), hypertrophic cardiomyopathy (HCM), cardiac amyloid (CA), and mitral valve prolapse (MVP). CNN-HMM derived ECG segmentation agreed with clinical estimates, with median absolute deviations (MAD) as a fraction of observed value of 0.6% for heart rate and 4% for QT interval. Patient-level ECG profiles enabled quantitative estimates of left ventricular and mitral annulus e' velocity with good discrimination in binary classification models of left ventricular hypertrophy and diastolic function. Models for disease detection ranged from AUROC of 0.94 to 0.77 for MVP. Top-ranked variables for all models included known ECG characteristics along with novel predictors of these traits/diseases.
A main puzzle of deep neural networks (DNNs) revolves around the apparent absence of "overfitting", defined in this paper as follows: the expected error does not get worse when increasing the number of neurons or of iterations of gradient descent. This is surprising because of the large capacity demonstrated by DNNs to fit randomly labeled data and the absence of explicit regularization. Recent results by Srebro et al. provide a satisfying solution of the puzzle for linear networks used in binary classification. They prove that minimization of loss functions such as the logistic, the cross-entropy and the exp-loss yields asymptotic, "slow" convergence to the maximum margin solution for linearly separable datasets, independently of the initial conditions. Here we prove a similar result for nonlinear multilayer DNNs near zero minima of the empirical loss. The result holds for exponential-type losses but not for the square loss. In particular, we prove that the weight matrix at each layer of a deep network converges to a minimum norm solution up to a scale factor (in the separable case). Our analysis of the dynamical system corresponding to gradient descent of a multilayer network suggests a simple criterion for ranking the generalization performance of different zero minimizers of the empirical loss.
Convolutional neural networks (CNN) have been shown to provide a good solution for classification problems that utilize data obtained from vibrational spectroscopy. Moreover, CNNs are capable of identification from noisy spectra without the need for additional preprocessing. However, their application in practical spectroscopy is limited due to two shortcomings. The effectiveness of the classification using CNNs drops rapidly when only a small number of spectra per substance are available for training (which is a typical situation in real applications). Additionally, to accommodate new, previously unseen substance classes, the network must be retrained which is computationally intensive. Here we address these issues by reformulating a multi-class classification problem with a large number of classes, but a small number of samples per class, to a binary classification problem with sufficient data available for representation learning. Namely, we define the learning task as identifying pairs of inputs as belonging to the same or different classes. We achieve this using a Siamese convolutional neural network. A novel sampling strategy is proposed to address the imbalance problem in training the Siamese Network. The trained network can effectively classify samples of unseen substance classes using just a single reference sample (termed as one-shot learning in the machine learning community). Our results demonstrate better accuracy than other practical systems to date, while allowing effortless updates of the system's database with novel substance classes.
We propose gradient adversarial training, an auxiliary deep learning framework applicable to different machine learning problems. In gradient adversarial training, we leverage a prior belief that in many contexts, simultaneous gradient updates should be statistically indistinguishable from each other. We enforce this consistency using an auxiliary network that classifies the origin of the gradient tensor, and the main network serves as an adversary to the auxiliary network in addition to performing standard task-based training. We demonstrate gradient adversarial training for three different scenarios: (1) as a defense to adversarial examples we classify gradient tensors and tune them to be agnostic to the class of their corresponding example, (2) for knowledge distillation, we do binary classification of gradient tensors derived from the student or teacher network and tune the student gradient tensor to mimic the teacher's gradient tensor; and (3) for multi-task learning we classify the gradient tensors derived from different task loss functions and tune them to be statistically indistinguishable. For each of the three scenarios we show the potential of gradient adversarial training procedure. Specifically, gradient adversarial training increases the robustness of a network to adversarial attacks, is able to better distill the knowledge from a teacher network to a student network compared to soft targets, and boosts multi-task learning by aligning the gradient tensors derived from the task specific loss functions. Overall, our experiments demonstrate that gradient tensors contain latent information about whatever tasks are being trained, and can support diverse machine learning problems when intelligently guided through adversarialization using a auxiliary network.
Fisher's linear discriminant analysis (FLDA) is an important dimension reduction method in statistical pattern recognition. It has been shown that FLDA is asymptotically Bayes optimal under the homoscedastic Gaussian assumption. However, this classical result has the following two major limitations: 1) it holds only for a fixed dimensionality $D$, and thus does not apply when $D$ and the training sample size $N$ are proportionally large; 2) it does not provide a quantitative description on how the generalization ability of FLDA is affected by $D$ and $N$. In this paper, we present an asymptotic generalization analysis of FLDA based on random matrix theory, in a setting where both $D$ and $N$ increase and $D/N\longrightarrow\gamma\in[0,1)$. The obtained lower bound of the generalization discrimination power overcomes both limitations of the classical result, i.e., it is applicable when $D$ and $N$ are proportionally large and provides a quantitative description of the generalization ability of FLDA in terms of the ratio $\gamma=D/N$ and the population discrimination power. Besides, the discrimination power bound also leads to an upper bound on the generalization error of binary-classification with FLDA.
Methods for analyzing or learning from "fuzzy data" have attracted increasing attention in recent years. In many cases, however, existing methods (for precise, non-fuzzy data) are extended to the fuzzy case in an ad-hoc manner, and without carefully considering the interpretation of a fuzzy set when being used for modeling data. Distinguishing between an ontic and an epistemic interpretation of fuzzy set-valued data, and focusing on the latter, we argue that a "fuzzification" of learning algorithms based on an application of the generic extension principle is not appropriate. In fact, the extension principle fails to properly exploit the inductive bias underlying statistical and machine learning methods, although this bias, at least in principle, offers a means for "disambiguating" the fuzzy data. Alternatively, we therefore propose a method which is based on the generalization of loss functions in empirical risk minimization, and which performs model identification and data disambiguation simultaneously. Elaborating on the fuzzification of specific types of losses, we establish connections to well-known loss functions in regression and classification. We compare our approach with related methods and illustrate its use in logistic regression for binary classification.
The bottom-up saliency, an early stage of humans' visual attention, can be considered as a binary classification problem between centre and surround classes. Discriminant power of features for the classification is measured as mutual information between distributions of image features and corresponding classes . As the estimated discrepancy very much depends on considered scale level, multi-scale structure and discriminant power are integrated by employing discrete wavelet features and Hidden Markov Tree (HMT). With wavelet coefficients and Hidden Markov Tree parameters, quad-tree like label structures are constructed and utilized in maximum a posterior probability (MAP) of hidden class variables at corresponding dyadic sub-squares. Then, a saliency value for each square block at each scale level is computed with discriminant power principle. Finally, across multiple scales is integrated the final saliency map by an information maximization rule. Both standard quantitative tools such as NSS, LCC, AUC and qualitative assessments are used for evaluating the proposed multi-scale discriminant saliency (MDIS) method against the well-know information based approach AIM on its released image collection with eye-tracking data. Simulation results are presented and analysed to verify the validity of MDIS as well as point out its limitation for further research direction.
We show how binary classification methods developed to work on i.i.d. data can be used for solving statistical problems that are seemingly unrelated to classification and concern highly-dependent time series. Specifically, the problems of time-series clustering, homogeneity testing and the three-sample problem are addressed. The algorithms that we construct for solving these problems are based on a new metric between time-series distributions, which can be evaluated using binary classification methods. Universal consistency of the proposed algorithms is proven under most general assumptions. The theoretical results are illustrated with experiments on synthetic and real-world data.
We introduce Supersparse Linear Integer Models (SLIM) as a tool to create scoring systems for binary classification. We derive theoretical bounds on the true risk of SLIM scoring systems, and present experimental results to show that SLIM scoring systems are accurate, sparse, and interpretable classification models.
The Gaussian Process Latent Variable Model (GP-LVM) is a non-linear probabilistic method of embedding a high dimensional dataset in terms low dimensional `latent' variables. In this paper we illustrate that maximum a posteriori (MAP) estimation of the latent variables and hyperparameters can be used for model selection and hence we can determine the optimal number or latent variables and the most appropriate model. This is an alternative to the variational approaches developed recently and may be useful when we want to use a non-Gaussian prior or kernel functions that don't have automatic relevance determination (ARD) parameters. Using a second order expansion of the latent variable posterior we can marginalise the latent variables and obtain an estimate for the hyperparameter posterior. Secondly, we use the GP-LVM to integrate multiple data sources by simultaneously embedding them in terms of common latent variables. We present results from synthetic data to illustrate the successful detection and retrieval of low dimensional structure from high dimensional data. We demonstrate that the integration of multiple data sources leads to more robust performance. Finally, we show that when the data are used for binary classification tasks we can attain a significant gain in prediction accuracy when the low dimensional representation is used.
In his seminal work, Schapire (1990) proved that weak classifiers could be improved to achieve arbitrarily high accuracy, but he never implied that a simple majority-vote mechanism could always do the trick. By comparing the asymptotic misclassification error of the majority-vote classifier with the average individual error, we discover an interesting phase-transition phenomenon. For binary classification with equal prior probabilities, our result implies that, for the majority-vote mechanism to work, the collection of weak classifiers must meet the minimum requirement of having an average true positive rate of at least 50% and an average false positive rate of at most 50%.
It has been found that stochastic algorithms often find good solutions much more rapidly than inherently-batch approaches. Indeed, a very useful rule of thumb is that often, when solving a machine learning problem, an iterative technique which relies on performing a very large number of relatively-inexpensive updates will often outperform one which performs a smaller number of much "smarter" but computationally-expensive updates. In this thesis, we will consider the application of stochastic algorithms to two of the most important machine learning problems. Part i is concerned with the supervised problem of binary classification using kernelized linear classifiers, for which the data have labels belonging to exactly two classes (e.g. "has cancer" or "doesn't have cancer"), and the learning problem is to find a linear classifier which is best at predicting the label. In Part ii, we will consider the unsupervised problem of Principal Component Analysis, for which the learning task is to find the directions which contain most of the variance of the data distribution. Our goal is to present stochastic algorithms for both problems which are, above all, practical--they work well on real-world data, in some cases better than all known competing algorithms. A secondary, but still very important, goal is to derive theoretical bounds on the performance of these algorithms which are at least competitive with, and often better than, those known for other approaches.
In this paper we propose a multi-armed bandit inspired, pool based active learning algorithm for the problem of binary classification. By carefully constructing an analogy between active learning and multi-armed bandits, we utilize ideas such as lower confidence bounds, and self-concordant regularization from the multi-armed bandit literature to design our proposed algorithm. Our algorithm is a sequential algorithm, which in each round assigns a sampling distribution on the pool, samples one point from this distribution, and queries the oracle for the label of this sampled point. The design of this sampling distribution is also inspired by the analogy between active learning and multi-armed bandits. We show how to derive lower confidence bounds required by our algorithm. Experimental comparisons to previously proposed active learning algorithms show superior performance on some standard UCI datasets.