We propose the task of automatically identifying papers used as baselines in a scientific article. We frame the problem as a binary classification task where all the references in a paper are to be classified as either baselines or non-baselines. This is a challenging problem due to the numerous ways in which a baseline reference can appear in a paper. We develop a dataset of $2,075$ papers from ACL anthology corpus with all their references manually annotated as one of the two classes. We develop a multi-module attention-based neural classifier for the baseline classification task that outperforms four state-of-the-art citation role classification methods when applied to the baseline classification task. We also present an analysis of the errors made by the proposed classifier, eliciting the challenges that make baseline identification a challenging problem.
Applying standard machine learning approaches for classification can produce unequal results across different demographic groups. When then used in real-world settings, these inequities can have negative societal impacts. This has motivated the development of various approaches to fair classification with machine learning models in recent years. In this paper, we consider the problem of modifying the predictions of a blackbox machine learning classifier in order to achieve fairness in a multiclass setting. To accomplish this, we extend the 'post-processing' approach in Hardt et al. 2016, which focuses on fairness for binary classification, to the setting of fair multiclass classification. We explore when our approach produces both fair and accurate predictions through systematic synthetic experiments and also evaluate discrimination-fairness tradeoffs on several publicly available real-world application datasets. We find that overall, our approach produces minor drops in accuracy and enforces fairness when the number of individuals in the dataset is high relative to the number of classes and protected groups.
Contrastive learning relies on an assumption that positive pairs contain related views, e.g., patches of an image or co-occurring multimodal signals of a video, that share certain underlying information about an instance. But what if this assumption is violated? The literature suggests that contrastive learning produces suboptimal representations in the presence of noisy views, e.g., false positive pairs with no apparent shared information. In this work, we propose a new contrastive loss function that is robust against noisy views. We provide rigorous theoretical justifications by showing connections to robust symmetric losses for noisy binary classification and by establishing a new contrastive bound for mutual information maximization based on the Wasserstein distance measure. The proposed loss is completely modality-agnostic and a simple drop-in replacement for the InfoNCE loss, which makes it easy to apply to existing contrastive frameworks. We show that our approach provides consistent improvements over the state-of-the-art on image, video, and graph contrastive learning benchmarks that exhibit a variety of real-world noise patterns.
The detection of early signs of volcanic unrest preceding an eruption, in the form of ground deformation in Interferometric Synthetic Aperture Radar (InSAR) data is critical for assessing volcanic hazard. In this work we treat this as a binary classification problem of InSAR images, and propose a novel deep learning methodology that exploits a rich source of synthetically generated interferograms to train quality classifiers that perform equally well in real interferograms. The imbalanced nature of the problem, with orders of magnitude fewer positive samples, coupled with the lack of a curated database with labeled InSAR data, sets a challenging task for conventional deep learning architectures. We propose a new framework for domain adaptation, in which we learn class prototypes from synthetic data with vision transformers. We report detection accuracy that amounts to the highest reported accuracy on a large test set for volcanic unrest detection. Moreover, we built upon this knowledge by learning a new, non-linear, projection between the learnt representations and prototype space, using pseudo labels produced by our model from an unlabeled real InSAR dataset. This leads to the new state of the art with 97.1% accuracy on our test set. We demonstrate the robustness of our approach by training a simple ResNet-18 Convolutional Neural Network on the unlabeled real InSAR dataset with pseudo-labels generated from our top transformer-prototype model. Our methodology provides a significant improvement in performance without the need of manually labeling any sample, opening the road for further exploitation of synthetic InSAR data in various remote sensing applications.
Knowledge graphs (KGs) are typically incomplete and we often wish to infer new facts given the existing ones. This can be thought of as a binary classification problem; we aim to predict if new facts are true or false. Unfortunately, we generally only have positive examples (the known facts) but we also need negative ones to train a classifier. To resolve this, it is usual to generate negative examples using a negative sampling strategy. However, this can produce false negatives which may reduce performance, is computationally expensive, and does not produce calibrated classification probabilities. In this paper, we propose a training procedure that obviates the need for negative sampling by adding a novel regularization term to the loss function. Our results for two relational embedding models (DistMult and SimplE) show the merit of our proposal both in terms of performance and speed.
Learning to optimize the area under the receiver operating characteristics curve (AUC) performance for imbalanced data has attracted much attention in recent years. Although there have been several methods of AUC optimization, scaling up AUC optimization is still an open issue due to its pairwise learning style. Maximizing AUC in the large-scale dataset can be considered as a non-convex and expensive problem. Inspired by the characteristic of pairwise learning, the cheap AUC optimization task with a small-scale dataset sampled from the large-scale dataset is constructed to promote the AUC accuracy of the original, large-scale, and expensive AUC optimization task. This paper develops an evolutionary multitasking framework (termed EMTAUC) to make full use of information among the constructed cheap and expensive tasks to obtain higher performance. In EMTAUC, one mission is to optimize AUC from the sampled dataset, and the other is to maximize AUC from the original dataset. Moreover, due to the cheap task containing limited knowledge, a strategy for dynamically adjusting the data structure of inexpensive tasks is proposed to introduce more knowledge into the multitasking AUC optimization environment. The performance of the proposed method is evaluated on a series of binary classification datasets. The experimental results demonstrate that EMTAUC is highly competitive to single task methods and online methods. Supplementary materials and source code implementation of EMTAUC can be accessed at https://github.com/xiaofangxd/EMTAUC.
The effect of bias on hypothesis formation is characterized for an automated data-driven projection pursuit neural network to extract and select features for binary classification of data streams. This intelligent exploratory process partitions a complete vector state space into disjoint subspaces to create working hypotheses quantified by similarities and differences observed between two groups of labeled data streams. Data streams are typically time sequenced, and may exhibit complex spatio-temporal patterns. For example, given atomic trajectories from molecular dynamics simulation, the machine's task is to quantify dynamical mechanisms that promote function by comparing protein mutants, some known to function while others are nonfunctional. Utilizing synthetic two-dimensional molecules that mimic the dynamics of functional and nonfunctional proteins, biases are identified and controlled in both the machine learning model and selected training data under different contexts. The refinement of a working hypothesis converges to a statistically robust multivariate perception of the data based on a context-dependent perspective. Including diverse perspectives during data exploration enhances interpretability of the multivariate characterization of similarities and differences.
There is an increasing interest in the application of deep learning architectures to tabular data. One of the state-of-the-art solutions is TabTransformer which incorporates an attention mechanism to better track relationships between categorical features and then makes use of a standard MLP to output its final logits. In this paper we propose multiple modifications to the original TabTransformer performing better on binary classification tasks for three separate datasets with more than 1% AUROC gains. Inspired by gated MLP, linear projections are implemented in the MLP block and multiple activation functions are tested. We also evaluate the importance of specific hyper parameters during training.
Federated learning allows clients to collaboratively learn statistical models while keeping their data local. Federated learning was originally used to train a unique global model to be served to all clients, but this approach might be sub-optimal when clients' local data distributions are heterogeneous. In order to tackle this limitation, recent personalized federated learning methods train a separate model for each client while still leveraging the knowledge available at other clients. In this work, we exploit the ability of deep neural networks to extract high quality vectorial representations (embeddings) from non-tabular data, e.g., images and text, to propose a personalization mechanism based on local memorization. Personalization is obtained by interpolating a collectively trained global model with a local $k$-nearest neighbors (kNN) model based on the shared representation provided by the global model. We provide generalization bounds for the proposed approach in the case of binary classification, and we show on a suite of federated datasets that this approach achieves significantly higher accuracy and fairness than state-of-the-art methods.
Deep models trained through maximum likelihood have achieved state-of-the-art results for survival analysis. Despite this training scheme, practitioners evaluate models under other criteria, such as binary classification losses at a chosen set of time horizons, e.g. Brier score (BS) and Bernoulli log likelihood (BLL). Models trained with maximum likelihood may have poor BS or BLL since maximum likelihood does not directly optimize these criteria. Directly optimizing criteria like BS requires inverse-weighting by the censoring distribution. However, estimating the censoring model under these metrics requires inverse-weighting by the failure distribution. The objective for each model requires the other, but neither are known. To resolve this dilemma, we introduce Inverse-Weighted Survival Games. In these games, objectives for each model are built from re-weighted estimates featuring the other model, where the latter is held fixed during training. When the loss is proper, we show that the games always have the true failure and censoring distributions as a stationary point. This means models in the game do not leave the correct distributions once reached. We construct one case where this stationary point is unique. We show that these games optimize BS on simulations and then apply these principles on real world cancer and critically-ill patient data.
Available data in machine learning applications is becoming increasingly complex, due to higher dimensionality and difficult classes. There exists a wide variety of approaches to measuring complexity of labeled data, according to class overlap, separability or boundary shapes, as well as group morphology. Many techniques can transform the data in order to find better features, but few focus on specifically reducing data complexity. Most data transformation methods mainly treat the dimensionality aspect, leaving aside the available information within class labels which can be useful when classes are somehow complex. This paper proposes an autoencoder-based approach to complexity reduction, using class labels in order to inform the loss function about the adequacy of the generated variables. This leads to three different new feature learners, Scorer, Skaler and Slicer. They are based on Fisher's discriminant ratio, the Kullback-Leibler divergence and least-squares support vector machines, respectively. They can be applied as a preprocessing stage for a binary classification problem. A thorough experimentation across a collection of 27 datasets and a range of complexity and classification metrics shows that class-informed autoencoders perform better than 4 other popular unsupervised feature extraction techniques, especially when the final objective is using the data for a classification task.
We consider the use of automated supervised learning systems for data tables that not only contain numeric/categorical columns, but one or more text fields as well. Here we assemble 18 multimodal data tables that each contain some text fields and stem from a real business application. Our publicly-available benchmark enables researchers to comprehensively evaluate their own methods for supervised learning with numeric, categorical, and text features. To ensure that any single modeling strategy which performs well over all 18 datasets will serve as a practical foundation for multimodal text/tabular AutoML, the diverse datasets in our benchmark vary greatly in: sample size, problem types (a mix of classification and regression tasks), number of features (with the number of text columns ranging from 1 to 28 between datasets), as well as how the predictive signal is decomposed between text vs. numeric/categorical features (and predictive interactions thereof). Over this benchmark, we evaluate various straightforward pipelines to model such data, including standard two-stage approaches where NLP is used to featurize the text such that AutoML for tabular data can then be applied. Compared with human data science teams, the fully automated methodology that performed best on our benchmark (stack ensembling a multimodal Transformer with various tree models) also manages to rank 1st place when fit to the raw text/tabular data in two MachineHack prediction competitions and 2nd place (out of 2380 teams) in Kaggle's Mercari Price Suggestion Challenge.
Over the past few years, several adversarial training methods have been proposed to improve the robustness of machine learning models against adversarial perturbations in the input. Despite remarkable progress in this regard, adversarial training is often observed to drop the standard test accuracy. This phenomenon has intrigued the research community to investigate the potential tradeoff between standard accuracy (a.k.a generalization) and robust accuracy (a.k.a robust generalization) as two performance measures. In this paper, we revisit this tradeoff for latent models and argue that this tradeoff is mitigated when the data enjoys a low-dimensional structure. In particular, we consider binary classification under two data generative models, namely Gaussian mixture model and generalized linear model, where the features data lie on a low-dimensional manifold. We develop a theory to show that the low-dimensional manifold structure allows one to obtain models that are nearly optimal with respect to both, the standard accuracy and the robust accuracy measures. We further corroborate our theory with several numerical experiments, including Mixture of Factor Analyzers (MFA) model trained on the MNIST dataset.
Genetic Programming (GP), an evolutionary learning technique, has multiple applications in machine learning such as curve fitting, data modelling, feature selection, classification etc. GP has several inherent parallel steps, making it an ideal candidate for GPU based parallelization. This paper describes a GPU accelerated stack-based variant of the generational GP algorithm which can be used for symbolic regression and binary classification. The selection and evaluation steps of the generational GP algorithm are parallelized using CUDA. We introduce representing candidate solution expressions as prefix lists, which enables evaluation using a fixed-length stack in GPU memory. CUDA based matrix vector operations are also used for computation of the fitness of population programs. We evaluate our algorithm on synthetic datasets for the Pagie Polynomial (ranging in size from $4096$ to $16$ million points), profiling training times of our algorithm with other standard symbolic regression libraries viz. gplearn, TensorGP and KarooGP. In addition, using $6$ large-scale regression and classification datasets usually used for comparing gradient boosting algorithms, we run performance benchmarks on our algorithm and gplearn, profiling the training time, test accuracy, and loss. On an NVIDIA DGX-A100 GPU, our algorithm outperforms all the previously listed frameworks, and in particular, achieves average speedups of $119\times$ and $40\times$ against gplearn on the synthetic and large scale datasets respectively.
Dynamic model pruning is a recent direction that allows for the inference of a different sub-network for each input sample during deployment. However, current dynamic methods rely on learning a continuous channel gating through regularization by inducing sparsity loss. This formulation introduces complexity in balancing different losses (e.g task loss, regularization loss). In addition, regularization based methods lack transparent tradeoff hyperparameter selection to realize a computational budget. Our contribution is two-fold: 1) decoupled task and pruning losses. 2) Simple hyperparameter selection that enables FLOPs reduction estimation before training. Inspired by the Hebbian theory in Neuroscience: "neurons that fire together wire together", we propose to predict a mask to process k filters in a layer based on the activation of its previous layer. We pose the problem as a self-supervised binary classification problem. Each mask predictor module is trained to predict if the log-likelihood for each filter in the current layer belongs to the top-k activated filters. The value k is dynamically estimated for each input based on a novel criterion using the mass of heatmaps. We show experiments on several neural architectures, such as VGG, ResNet and MobileNet on CIFAR and ImageNet datasets. On CIFAR, we reach similar accuracy to SOTA methods with 15% and 24% higher FLOPs reduction. Similarly in ImageNet, we achieve lower drop in accuracy with up to 13% improvement in FLOPs reduction.
Despite the pervasiveness of ordinal labels in supervised learning, it remains common practice in deep learning to treat such problems as categorical classification using the categorical cross entropy loss. Recent methods attempting to address this issue while respecting the ordinal structure of the labels have resorted to converting ordinal regression into a series of extended binary classification subtasks. However, the adoption of such methods remains inconsistent due to theoretical and practical limitations. Here we address these limitations by demonstrating that the subtask probabilities form a Markov chain. We show how to straightforwardly modify neural network architectures to exploit this fact and thereby constrain predictions to be universally rank consistent. We furthermore prove that all rank consistent solutions can be represented within this formulation, and derive a loss function producing maximum likelihood parameter estimates. Using diverse benchmarks and the real-world application of a specialized recurrent neural network for COVID-19 prognosis, we demonstrate the practical superiority of this method versus the current state-of-the-art. The method is open sourced as user-friendly PyTorch and TensorFlow packages.
Group equivariant convolutional neural networks (G-CNNs) are generalizations of convolutional neural networks (CNNs) which excel in a wide range of technical applications by explicitly encoding symmetries, such as rotations and permutations, in their architectures. Although the success of G-CNNs is driven by their \emph{explicit} symmetry bias, a recent line of work has proposed that the \emph{implicit} bias of training algorithms on particular architectures is key to understanding generalization for overparameterized neural nets. In this context, we show that $L$-layer full-width linear G-CNNs trained via gradient descent for binary classification converge to solutions with low-rank Fourier matrix coefficients, regularized by the $2/L$-Schatten matrix norm. Our work strictly generalizes previous analysis on the implicit bias of linear CNNs to linear G-CNNs over all finite groups, including the challenging setting of non-commutative groups (such as permutations), as well as band-limited G-CNNs over infinite groups. We validate our theorems via experiments on a variety of groups, and empirically explore more realistic nonlinear networks, which locally capture similar regularization patterns. Finally, we provide intuitive interpretations of our Fourier space implicit regularization results in real space via uncertainty principles.
Contrastive learning has demonstrated promising performance in image and text domains either in a self-supervised or a supervised manner. In this work, we extend the supervised contrastive learning framework to clinical risk prediction problems based on longitudinal electronic health records (EHR). We propose a general supervised contrastive loss $\mathcal{L}_{\text{Contrastive Cross Entropy} } + \lambda \mathcal{L}_{\text{Supervised Contrastive Regularizer}}$ for learning both binary classification (e.g. in-hospital mortality prediction) and multi-label classification (e.g. phenotyping) in a unified framework. Our supervised contrastive loss practices the key idea of contrastive learning, namely, pulling similar samples closer and pushing dissimilar ones apart from each other, simultaneously by its two components: $\mathcal{L}_{\text{Contrastive Cross Entropy} }$ tries to contrast samples with learned anchors which represent positive and negative clusters, and $\mathcal{L}_{\text{Supervised Contrastive Regularizer}}$ tries to contrast samples with each other according to their supervised labels. We propose two versions of the above supervised contrastive loss and our experiments on real-world EHR data demonstrate that our proposed loss functions show benefits in improving the performance of strong baselines and even state-of-the-art models on benchmarking tasks for clinical risk predictions. Our loss functions work well with extremely imbalanced data which are common for clinical risk prediction problems. Our loss functions can be easily used to replace (binary or multi-label) cross-entropy loss adopted in existing clinical predictive models. The Pytorch code is released at \url{https://github.com/calvin-zcx/SCEHR}.
In this paper, we present FoodChem, a new Relation Extraction (RE) model for identifying chemicals present in the composition of food entities, based on textual information provided in biomedical peer-reviewed scientific literature. The RE task is treated as a binary classification problem, aimed at identifying whether the contains relation exists between a food-chemical entity pair. This is accomplished by fine-tuning BERT, BioBERT and RoBERTa transformer models. For evaluation purposes, a novel dataset with annotated contains relations in food-chemical entity pairs is generated, in a golden and silver version. The models are integrated into a voting scheme in order to produce the silver version of the dataset which we use for augmenting the individual models, while the manually annotated golden version is used for their evaluation. Out of the three evaluated models, the BioBERT model achieves the best results, with a macro averaged F1 score of 0.902 in the unbalanced augmentation setting.
As online news has become increasingly popular and fake news increasingly prevalent, the ability to audit the veracity of online news content has become more important than ever. Such a task represents a binary classification challenge, for which transformers have achieved state-of-the-art results. Using the publicly available ISOT and Combined Corpus datasets, this study explores transformers' abilities to identify fake news, with particular attention given to investigating generalisation to unseen datasets with varying styles, topics and class distributions. Moreover, we explore the idea that opinion-based news articles cannot be classified as real or fake due to their subjective nature and often sensationalised language, and propose a novel two-step classification pipeline to remove such articles from both model training and the final deployed inference system. Experiments over the ISOT and Combined Corpus datasets show that transformers achieve an increase in F1 scores of up to 4.9% for out of distribution generalisation compared to baseline approaches, with a further increase of 10.1% following the implementation of our two-step classification pipeline. To the best of our knowledge, this study is the first to investigate generalisation of transformers in this context.