What is the value of an individual model in an ensemble of binary classifiers? We answer this question by introducing a class of transferable utility cooperative games called \textit{ensemble games}. In machine learning ensembles, pre-trained models cooperate to make classification decisions. To quantify the importance of models in these ensemble games, we define \textit{Troupe} -- an efficient algorithm which allocates payoffs based on approximate Shapley values of the classifiers. We argue that the Shapley value of models in these games is an effective decision metric for choosing a high performing subset of models from the ensemble. Our analytical findings prove that our Shapley value estimation scheme is precise and scalable; its performance increases with size of the dataset and ensemble. Empirical results on real world graph classification tasks demonstrate that our algorithm produces high quality estimates of the Shapley value. We find that Shapley values can be utilized for ensemble pruning, and that adversarial models receive a low valuation. Complex classifiers are frequently found to be responsible for both correct and incorrect classification decisions.
As 3D point cloud analysis has received increasing attention, the insufficient scale of point cloud datasets and the weak generalization ability of networks become prominent. In this paper, we propose a simple and effective augmentation method for the point cloud data, named PointCutMix, to alleviate those problems. It finds the optimal assignment between two point clouds and generates new training data by replacing the points in one sample with their optimal assigned pairs. Two replacement strategies are proposed to adapt to the accuracy or robustness requirement for different tasks, one of which is to randomly select all replacing points while the other one is to select k nearest neighbors of a single random point. Both strategies consistently and significantly improve the performance of various models on point cloud classification problems. By introducing the saliency maps to guide the selection of replacing points, the performance further improves. Moreover, PointCutMix is validated to enhance the model robustness against the point attack. It is worth noting that when using as a defense method, our method outperforms the state-of-the-art defense algorithms. The code is available at:https://github.com/cuge1995/PointCutMix
Joint Energy-based Model (JEM) of Grathwohl et al. shows that a standard softmax classifier can be reinterpreted as an energy-based model (EBM) for the joint distribution p(x,y); the resulting model can be optimized to improve calibration, robustness, and out-of-distribution detection, while generating samples rivaling the quality of recent GAN-based approaches. However, the softmax classifier that JEM exploits is inherently discriminative and its latent feature space is not well formulated as probabilistic distributions, which may hinder its potential for image generation and incur training instability. We hypothesize that generative classifiers, such as Linear Discriminant Analysis (LDA), might be more suitable for image generation since generative classifiers model the data generation process explicitly. This paper therefore investigates an LDA classifier for image classification and generation. In particular, the Max-Mahalanobis Classifier (MMC), a special case of LDA, fits our goal very well. We show that our Generative MMC (GMMC) can be trained discriminatively, generatively, or jointly for image classification and generation. Extensive experiments on multiple datasets show that GMMC achieves state-of-the-art discriminative and generative performances, while outperforming JEM in calibration, adversarial robustness, and out-of-distribution detection by a significant margin. Our source code is available at https://github.com/sndnyang/GMMC.
Recent papers have suggested that transfer learning can outperform sophisticated meta-learning methods for few-shot image classification. We take this hypothesis to its logical conclusion, and suggest the use of an ensemble of high-quality, pre-trained feature extractors for few-shot image classification. We show experimentally that a library of pre-trained feature extractors combined with a simple feed-forward network learned with an L2-regularizer can be an excellent option for solving cross-domain few-shot image classification. Our experimental results suggest that this simpler sample-efficient approach far outperforms several well-established meta-learning algorithms on a variety of few-shot tasks.
We study the following problem: given two public transit station identifiers A and B, each with a label and a geographic coordinate, decide whether A and B describe the same station. For example, for "St Pancras International" at (51.5306, -0.1253) and "London St Pancras" at (51.5319, -0.1269), the answer would be "Yes". This problem frequently arises in areas where public transit data is used, for example in geographic information systems, schedule merging, route planning, or map matching. We consider several baseline methods based on geographic distance and simple string similarity measures. We also experiment with more elaborate string similarity measures and manually created normalization rules. Our experiments show that these baseline methods produce good, but not fully satisfactory results. We therefore develop an approach based on a random forest classifier which is trained on matching trigrams between two stations, their distance, and their position on an interwoven grid. All approaches are evaluated on extensive ground truth datasets we generated from OpenStreetMap (OSM) data: (1) The union of Great Britain and Ireland and (2) the union of Germany, Switzerland, and Austria. On all datasets, our learning-based approach achieves an F1 score of over 99%, while even the most elaborate baseline approach (based on TFIDF scores and the geographic distance) achieves an F1 score of at most 94%, and a naive approach of using a geographical distance threshold achieves an F1 score of only 75%. Both our training and testing datasets are publicly available.
Feature representation is an important aspect of remote-sensing based image classification. While deep convolutional neural networks are able to effectively amalgamate information, large numbers of parameters often make learned features inscrutable and difficult to transfer to alternative models. In order to better represent statistical texture information for remote-sensing image classification, in this paper, we investigate performing joint dimensionality reduction and classification using a novel histogram neural network. Motivated by a popular dimensionality reduction approach, t-Distributed Stochastic Neighbor Embedding (t-SNE), our proposed method incorporates a classification loss computed on samples in a low-dimensional embedding space. We compare the learned sample embeddings against coordinates found by t-SNE in terms of classification accuracy and qualitative assessment. We also explore use of various divergence measures in the t-SNE objective. The proposed method has several advantages such as readily embedding out-of-sample points and reducing feature dimensionality while retaining class discriminability. Our results show that the proposed approach maintains and/or improves classification performance and reveals characteristics of features produced by neural networks that may be helpful for other applications.
Semi-supervised learning has been gaining attention as it allows for performing image analysis tasks such as classification with limited labeled data. Some popular algorithms using Generative Adversarial Networks (GANs) for semi-supervised classification share a single architecture for classification and discrimination. However, this may require a model to converge to a separate data distribution for each task, which may reduce overall performance. While progress in semi-supervised learning has been made, less addressed are small-scale, fully-supervised tasks where even unlabeled data is unavailable and unattainable. We therefore, propose a novel GAN model namely External Classifier GAN (EC-GAN), that utilizes GANs and semi-supervised algorithms to improve classification in fully-supervised regimes. Our method leverages a GAN to generate artificial data used to supplement supervised classification. More specifically, we attach an external classifier, hence the name EC-GAN, to the GAN's generator, as opposed to sharing an architecture with the discriminator. Our experiments demonstrate that EC-GAN's performance is comparable to the shared architecture method, far superior to the standard data augmentation and regularization-based approach, and effective on a small, realistic dataset.
In this paper, we explore contrastive learning for few-shot classification, in which we propose to use it as an additional auxiliary training objective acting as a data-dependent regularizer to promote more general and transferable features. In particular, we present a novel attention-based spatial contrastive objective to learn locally discriminative and class-agnostic features. As a result, our approach overcomes some of the limitations of the cross-entropy loss, such as its excessive discrimination towards seen classes, which reduces the transferability of features to unseen classes. With extensive experiments, we show that the proposed method outperforms state-of-the-art approaches, confirming the importance of learning good and transferable embeddings for few-shot learning.
Choosing a decision threshold is one of the challenging job in any classification tasks. How much the model is accurate, if the deciding boundary is not picked up carefully, its entire performance would go in vain. On the other hand, for imbalance classification where one of the classes is dominant over another, relying on the conventional method of choosing threshold would result in poor performance. Even if the threshold or decision boundary is properly chosen based on machine learning strategies like SVM and decision tree, it will fail at some point for dynamically varying databases and in case of identity-features that are more or less similar, like in face recognition and person re-identification models. Hence, with the need for adaptability of the decision threshold selection for imbalanced classification and incremental database size, an online optimization-based statistical feature learning adaptive technique is developed and tested on the LFW datasets and self-prepared athletes datasets. This method of adopting adaptive threshold resulted in 12-45% improvement in the model accuracy compared to the fixed threshold {0.3,0.5,0.7} that are usually taken via the hit-and-trial method in any classification and identification tasks. Source code for the complete algorithm is available at: https://github.com/Varat7v2/adaptive-threshold
We combine K-means clustering with the least-squares kernel classification method. K-means clustering is used to extract a set of representative vectors for each class. The least-squares kernel method uses these representative vectors as a training set for the classification task. We show that this combination of unsupervised and supervised learning algorithms performs very well, and we illustrate this approach using the MNIST dataset
Compared with single-label image classification, multi-label image classification is more practical and challenging. Some recent studies attempted to leverage the semantic information of categories for improving multi-label image classification performance. However, these semantic-based methods only take semantic information as type of complements for visual representation without further exploitation. In this paper, we present an innovative path towards the solution of the multi-label image classification which considers it as a dictionary learning task. A novel end-to-end model named Deep Semantic Dictionary Learning (DSDL) is designed. In DSDL, an auto-encoder is applied to generate the semantic dictionary from class-level semantics and then such dictionary is utilized for representing the visual features extracted by Convolutional Neural Network (CNN) with label embeddings. The DSDL provides a simple but elegant way to exploit and reconcile the label, semantic and visual spaces simultaneously via conducting the dictionary learning among them. Moreover, inspired by iterative optimization of traditional dictionary learning, we further devise a novel training strategy named Alternately Parameters Update Strategy (APUS) for optimizing DSDL, which alternately optimizes the representation coefficients and the semantic dictionary in forward and backward propagation. Extensive experimental results on three popular benchmarks demonstrate that our method achieves promising performances in comparison with the state-of-the-arts. Our codes and models have been released at {https://github.com/ZFT-CQU/DSDL}.
Motor imagery brain computer interface designs are considered difficult due to limitations in subject-specific data collection and calibration, as well as demanding system adaptation requirements. Recently, subject-independent (SI) designs received attention because of their possible applicability to multiple users without prior calibration and rigorous system adaptation. SI designs are challenging and have shown low accuracy in the literature. Two major factors in system performance are the classification algorithm and the quality of available data. This paper presents a comparative study of classification performance for both SS and SI paradigms. Our results show that classification algorithms for SS models display large variance in performance. Therefore, distinct classification algorithms per subject may be required. SI models display lower variance in performance but should only be used if a relatively large sample size is available. For SI models, LDA and CART had the highest accuracy for small and moderate sample size, respectively, whereas we hypothesize that SVM would be superior to the other classifiers if large training sample-size was available. Additionally, one should choose the design approach considering the users. While the SS design sound more promising for a specific subject, an SI approach can be more convenient for mentally or physically challenged users.
Anomalies are ubiquitous in all scientific fields and can express an unexpected event due to incomplete knowledge about the data distribution or an unknown process that suddenly comes into play and distorts observations. Due to such events' rarity, to train deep learning models on the Anomaly Detection (AD) task, scientists only rely on "normal" data, i.e., non-anomalous samples. Thus, letting the neural network infer the distribution beneath the input data. In such a context, we propose a novel framework, named Multi-layer One-Class ClassificAtion (MOCCA),to train and test deep learning models on the AD task. Specifically, we applied it to autoencoders. A key novelty in our work stems from the explicit optimization of intermediate representations for the AD task. Indeed, differently from commonly used approaches that consider a neural network as a single computational block, i.e., using the output of the last layer only, MOCCA explicitly leverages the multi-layer structure of deep architectures. Each layer's feature space is optimized for AD during training, while in the test phase, the deep representations extracted from the trained layers are combined to detect anomalies. With MOCCA, we split the training process into two steps. First, the autoencoder is trained on the reconstruction task only. Then, we only retain the encoder tasked with minimizing the L_2 distance between the output representation and a reference point, the anomaly-free training data centroid, at each considered layer. Subsequently, we combine the deep features extracted at the various trained layers of the encoder model to detect anomalies at inference time. To assess the performance of the models trained with MOCCA, we conduct extensive experiments on publicly available datasets. We show that our proposed method reaches comparable or superior performance to state-of-the-art approaches available in the literature.
Convolutional Networks have dominated the field of computer vision for the last ten years, exhibiting extremely powerful feature extraction capabilities and outstanding classification performance. The main strategy to prolong this trend relies on further upscaling networks in size. However, costs increase rapidly while performance improvements may be marginal. We hypothesise that adding heterogeneous sources of information may be more cost-effective to a CNN than building a bigger network. In this paper, an ensemble method is proposed for accurate image classification, fusing automatically detected features through Convolutional Neural Network architectures with a set of manually defined statistical indicators. Through a combination of the predictions of a CNN and a secondary classifier trained on statistical features, better classification performance can be cheaply achieved. We test multiple learning algorithms and CNN architectures on a diverse number of datasets to validate our proposal, making public all our code and data via GitHub. According to our results, the inclusion of additional indicators and an ensemble classification approach helps to increase the performance in 8 of 9 datasets, with a remarkable increase of more than 10% precision in two of them.
In this paper, we propose a novel method named Biomedical Confident Itemsets Explanation (BioCIE), aiming at post-hoc explanation of black-box machine learning models for biomedical text classification. Using sources of domain knowledge and a confident itemset mining method, BioCIE discretizes the decision space of a black-box into smaller subspaces and extracts semantic relationships between the input text and class labels in different subspaces. Confident itemsets discover how biomedical concepts are related to class labels in the black-box's decision space. BioCIE uses the itemsets to approximate the black-box's behavior for individual predictions. Optimizing fidelity, interpretability, and coverage measures, BioCIE produces class-wise explanations that represent decision boundaries of the black-box. Results of evaluations on various biomedical text classification tasks and black-box models demonstrated that BioCIE can outperform perturbation-based and decision set methods in terms of producing concise, accurate, and interpretable explanations. BioCIE improved the fidelity of instance-wise and class-wise explanations by 11.6% and 7.5%, respectively. It also improved the interpretability of explanations by 8%. BioCIE can be effectively used to explain how a black-box biomedical text classification model semantically relates input texts to class labels. The source code and supplementary material are available at https://github.com/mmoradi-iut/BioCIE.
Machine learning techniques can be useful in applications such as credit approval and college admission. However, to be classified more favorably in such contexts, an agent may decide to strategically withhold some of her features, such as bad test scores. This is a missing data problem with a twist: which data is missing {\em depends on the chosen classifier}, because the specific classifier is what may create the incentive to withhold certain feature values. We address the problem of training classifiers that are robust to this behavior. We design three classification methods: {\sc Mincut}, {\sc Hill-Climbing} ({\sc HC}) and Incentive-Compatible Logistic Regression ({\sc IC-LR}). We show that {\sc Mincut} is optimal when the true distribution of data is fully known. However, it can produce complex decision boundaries, and hence be prone to overfitting in some cases. Based on a characterization of truthful classifiers (i.e., those that give no incentive to strategically hide features), we devise a simpler alternative called {\sc HC} which consists of a hierarchical ensemble of out-of-the-box classifiers, trained using a specialized hill-climbing procedure which we show to be convergent. For several reasons, {\sc Mincut} and {\sc HC} are not effective in utilizing a large number of complementarily informative features. To this end, we present {\sc IC-LR}, a modification of Logistic Regression that removes the incentive to strategically drop features. We also show that our algorithms perform well in experiments on real-world data sets, and present insights into their relative performance in different settings.
Open intent classification is a challenging task in dialogue systems. On the one hand, it should ensure the quality of known intent identification. On the other hand, it needs to detect the open (unknown) intent without prior knowledge. Current models are limited in finding the appropriate decision boundary to balance the performances of both known intents and the open intent. In this paper, we propose a post-processing method to learn the adaptive decision boundary (ADB) for open intent classification. We first utilize the labeled known intent samples to pre-train the model. Then, we automatically learn the adaptive spherical decision boundary for each known class with the aid of well-trained features. Specifically, we propose a new loss function to balance both the empirical risk and the open space risk. Our method does not need open intent samples and is free from modifying the model architecture. Moreover, our approach is surprisingly insensitive with less labeled data and fewer known intents. Extensive experiments on three benchmark datasets show that our method yields significant improvements compared with the state-of-the-art methods. The codes are released at https://github.com/thuiar/Adaptive-Decision-Boundary.
Self-attention techniques, and specifically Transformers, are dominating the field of text processing and are becoming increasingly popular in computer vision classification tasks. In order to visualize the parts of the image that led to a certain classification, existing methods either rely on the obtained attention maps or employ heuristic propagation along the attention graph. In this work, we propose a novel way to compute relevancy for Transformer networks. The method assigns local relevance based on the Deep Taylor Decomposition principle and then propagates these relevancy scores through the layers. This propagation involves attention layers and skip connections, which challenge existing methods. Our solution is based on a specific formulation that is shown to maintain the total relevancy across layers. We benchmark our method on very recent visual Transformer networks, as well as on a text classification problem, and demonstrate a clear advantage over the existing explainability methods.
Pre-trained language models have achieved state-of-the-art accuracies on various text classification tasks, e.g., sentiment analysis, natural language inference, and semantic textual similarity. However, the reliability of the fine-tuned text classifiers is an often underlooked performance criterion. For instance, one may desire a model that can detect out-of-distribution (OOD) samples (drawn far from training distribution) or be robust against domain shifts. We claim that one central obstacle to the reliability is the over-reliance of the model on a limited number of keywords, instead of looking at the whole context. In particular, we find that (a) OOD samples often contain in-distribution keywords, while (b) cross-domain samples may not always contain keywords; over-relying on the keywords can be problematic for both cases. In light of this observation, we propose a simple yet effective fine-tuning method, coined masked keyword regularization (MASKER), that facilitates context-based prediction. MASKER regularizes the model to reconstruct the keywords from the rest of the words and make low-confidence predictions without enough context. When applied to various pre-trained language models (e.g., BERT, RoBERTa, and ALBERT), we demonstrate that MASKER improves OOD detection and cross-domain generalization without degrading classification accuracy. Code is available at https://github.com/alinlab/MASKER.
Hate speech, quite common in the age of social media, at times harmless but can also cause mental trauma to someone or even riots in communities. Image of a religious symbol with derogatory comment or video of a man abusing a particular community, all become hate speech with its every modality (such as text, image, and audio) contributing towards it. Models based on a particular modality of hate speech post on social media are not useful, rather, we need models like multi-modal fusion models that consider both image and text while classifying hate speech. Text-image fusion models are heavily parameterized, hence we propose a quaternion neural network-based model having additional fusion components for each pair of modalities. The model is tested on the MMHS150K twitter dataset for hate speech classification. The model shows an almost 75% reduction in parameters and also benefits us in terms of storage space and training time while being at par in terms of performance as compared to its real counterpart.