Progress on many Natural Language Processing (NLP) tasks, such as text classification, is driven by objective, reproducible and scalable evaluation via publicly available benchmarks. However, these are not always representative of real-world scenarios where text classifiers are employed, such as sentiment analysis or misinformation detection. In this position paper, we put forward two points that aim to alleviate this problem. First, we propose to extend text classification benchmarks to evaluate the explainability of text classifiers. We review challenges associated with objectively evaluating the capabilities to produce valid explanations which leads us to the second main point: We propose to ground these benchmarks in human-centred applications, for example by using social media, gamification or to learn explainability metrics from human judgements.
Machine learning and deep learning classification models are data-driven, and the model and the data jointly determine their classification performance. It is biased to evaluate the model's performance only based on the classifier accuracy while ignoring the data separability. Sometimes, the model exhibits excellent accuracy, which might be attributed to its testing on highly separable data. Most of the current studies on data separability measures are defined based on the distance between sample points, but this has been demonstrated to fail in several circumstances. In this paper, we propose a new separability measure--the rate of separability (RS), which is based on the data coding rate. We validate its effectiveness as a supplement to the separability measure by comparing it to four other distance-based measures on synthetic datasets. Then, we demonstrate the positive correlation between the proposed measure and recognition accuracy in a multi-task scenario constructed from a real dataset. Finally, we discuss the methods for evaluating the classification performance of machine learning and deep learning models considering data separability.
Metric-based meta-learning is one of the de facto standards in few-shot learning. It composes of representation learning and metrics calculation designs. Previous works construct class representations in different ways, varying from mean output embedding to covariance and distributions. However, using embeddings in space lacks expressivity and cannot capture class information robustly, while statistical complex modeling poses difficulty to metric designs. In this work, we use tensor fields (``areas'') to model classes from the geometrical perspective for few-shot learning. We present a simple and effective method, dubbed hypersphere prototypes (HyperProto), where class information is represented by hyperspheres with dynamic sizes with two sets of learnable parameters: the hypersphere's center and the radius. Extending from points to areas, hyperspheres are much more expressive than embeddings. Moreover, it is more convenient to perform metric-based classification with hypersphere prototypes than statistical modeling, as we only need to calculate the distance from a data point to the surface of the hypersphere. Following this idea, we also develop two variants of prototypes under other measurements. Extensive experiments and analysis on few-shot learning tasks across NLP and CV and comparison with 20+ competitive baselines demonstrate the effectiveness of our approach.
Compared to the great progress of large-scale vision transformers (ViTs) in recent years, large-scale models based on convolutional neural networks (CNNs) are still in an early state. This work presents a new large-scale CNN-based foundation model, termed InternImage, which can obtain the gain from increasing parameters and training data like ViTs. Different from the recent CNNs that focus on large dense kernels, InternImage takes deformable convolution as the core operator, so that our model not only has the large effective receptive field required for downstream tasks such as detection and segmentation, but also has the adaptive spatial aggregation conditioned by input and task information. As a result, the proposed InternImage reduces the strict inductive bias of traditional CNNs and makes it possible to learn stronger and more robust patterns with large-scale parameters from massive data like ViTs. The effectiveness of our model is proven on challenging benchmarks including ImageNet, COCO, and ADE20K. It is worth mentioning that InternImage-H achieved a new record 65.4 mAP on COCO test-dev and 62.9 mIoU on ADE20K, outperforming current leading CNNs and ViTs. The code will be released at https://github.com/OpenGVLab/InternImage.
As the Internet grows in popularity, more and more classification jobs, such as IoT, finance industry and healthcare field, rely on mobile edge computing to advance machine learning. In the medical industry, however, good diagnostic accuracy necessitates the combination of large amounts of labeled data to train the model, which is difficult and expensive to collect and risks jeopardizing patients' privacy. In this paper, we offer a novel medical diagnostic framework that employs a federated learning platform to ensure patient data privacy by transferring classification algorithms acquired in a labeled domain to a domain with sparse or missing labeled data. Rather than using a generative adversarial network, our framework uses a discriminative model to build multiple classification loss functions with the goal of improving diagnostic accuracy. It also avoids the difficulty of collecting large amounts of labeled data or the high cost of generating large amount of sample data. Experiments on real-world image datasets demonstrates that the suggested adversarial federated transfer learning method is promising for real-world medical diagnosis applications that use image classification.
Modern neural networks are over-parameterized and thus rely on strong regularization such as data augmentation and weight decay to reduce overfitting and improve generalization. The dominant form of data augmentation applies invariant transforms, where the learning target of a sample is invariant to the transform applied to that sample. We draw inspiration from human visual classification studies and propose generalizing augmentation with invariant transforms to soft augmentation where the learning target softens non-linearly as a function of the degree of the transform applied to the sample: e.g., more aggressive image crop augmentations produce less confident learning targets. We demonstrate that soft targets allow for more aggressive data augmentation, offer more robust performance boosts, work with other augmentation policies, and interestingly, produce better calibrated models (since they are trained to be less confident on aggressively cropped/occluded examples). Combined with existing aggressive augmentation strategies, soft target 1) doubles the top-1 accuracy boost across Cifar-10, Cifar-100, ImageNet-1K, and ImageNet-V2, 2) improves model occlusion performance by up to $4\times$, and 3) halves the expected calibration error (ECE). Finally, we show that soft augmentation generalizes to self-supervised classification tasks. Code available at https://github.com/youngleox/soft_augmentation
Spiking Neural Networks are often touted as brain-inspired learning models for the third wave of Artificial Intelligence. Although recent SNNs trained with supervised backpropagation show classification accuracy comparable to deep networks, the performance of unsupervised learning-based SNNs remains much lower. This paper presents a heterogeneous recurrent spiking neural network (HRSNN) with unsupervised learning for spatio-temporal classification of video activity recognition tasks on RGB (KTH, UCF11, UCF101) and event-based datasets (DVS128 Gesture). The key novelty of the HRSNN is that the recurrent layer in HRSNN consists of heterogeneous neurons with varying firing/relaxation dynamics, and they are trained via heterogeneous spike-time-dependent-plasticity (STDP) with varying learning dynamics for each synapse. We show that this novel combination of heterogeneity in architecture and learning method outperforms current homogeneous spiking neural networks. We further show that HRSNN can achieve similar performance to state-of-the-art backpropagation trained supervised SNN, but with less computation (fewer neurons and sparse connection) and less training data.
A new development in NLP is the construction of hyperbolic word embeddings. As opposed to their Euclidean counterparts, hyperbolic embeddings are represented not by vectors, but by points in hyperbolic space. This makes the most common basic scheme for constructing document representations, namely the averaging of word vectors, meaningless in the hyperbolic setting. We reinterpret the vector mean as the centroid of the points represented by the vectors, and investigate various hyperbolic centroid schemes and their effectiveness at text classification.
In this paper, we introduced the novel concept of advisor network to address the problem of noisy labels in image classification. Deep neural networks (DNN) are prone to performance reduction and overfitting problems on training data with noisy annotations. Weighting loss methods aim to mitigate the influence of noisy labels during the training, completely removing their contribution. This discarding process prevents DNNs from learning wrong associations between images and their correct labels but reduces the amount of data used, especially when most of the samples have noisy labels. Differently, our method weighs the feature extracted directly from the classifier without altering the loss value of each data. The advisor helps to focus only on some part of the information present in mislabeled examples, allowing the classifier to leverage that data as well. We trained it with a meta-learning strategy so that it can adapt throughout the training of the main model. We tested our method on CIFAR10 and CIFAR100 with synthetic noise, and on Clothing1M which contains real-world noise, reporting state-of-the-art results.
Humans can classify data of an unseen category by reasoning on its language explanations. This ability is owing to the compositional nature of language: we can combine previously seen attributes to describe the new category. For example, we might describe a sage thrasher as "it has a slim straight relatively short bill, yellow eyes and a long tail", so that others can use their knowledge of attributes "slim straight relatively short bill", "yellow eyes" and "long tail" to recognize a sage thrasher. Inspired by this observation, in this work we tackle zero-shot classification task by logically parsing and reasoning on natural language expla-nations. To this end, we propose the framework CLORE (Classification by LOgical Reasoning on Explanations). While previous methods usually regard textual information as implicit features, CLORE parses explanations into logical structures and then explicitly reasons along thess structures on the input to produce a classification score. Experimental results on explanation-based zero-shot classification benchmarks demonstrate that CLORE is superior to baselines, which we further show mainly comes from higher scores on tasks requiring more logical reasoning. We also demonstrate that our framework can be extended to zero-shot classification on visual modality. Alongside classification decisions, CLORE can provide the logical parsing and reasoning process as a clear form of rationale. Through empirical analysis we demonstrate that CLORE is also less affected by linguistic biases than baselines.
Automatic topic classification has been studied extensively to assist managing and indexing scientific documents in a digital collection. With the large number of topics being available in recent years, it has become necessary to arrange them in a hierarchy. Therefore, the automatic classification systems need to be able to classify the documents hierarchically. In addition, each paper is often assigned to more than one relevant topic. For example, a paper can be assigned to several topics in a hierarchy tree. In this paper, we introduce a new dataset for hierarchical multi-label text classification (HMLTC) of scientific papers called SciHTC, which contains 186,160 papers and 1,233 categories from the ACM CCS tree. We establish strong baselines for HMLTC and propose a multi-task learning approach for topic classification with keyword labeling as an auxiliary task. Our best model achieves a Macro-F1 score of 34.57% which shows that this dataset provides significant research opportunities on hierarchical scientific topic classification. We make our dataset and code available on Github.
Graph classification is an important area in both modern research and industry. Multiple applications, especially in chemistry and novel drug discovery, encourage rapid development of machine learning models in this area. To keep up with the pace of new research, proper experimental design, fair evaluation, and independent benchmarks are essential. Design of strong baselines is an indispensable element of such works. In this thesis, we explore multiple approaches to graph classification. We focus on Graph Neural Networks (GNNs), which emerged as a de facto standard deep learning technique for graph representation learning. Classical approaches, such as graph descriptors and molecular fingerprints, are also addressed. We design fair evaluation experimental protocol and choose proper datasets collection. This allows us to perform numerous experiments and rigorously analyze modern approaches. We arrive to many conclusions, which shed new light on performance and quality of novel algorithms. We investigate application of Jumping Knowledge GNN architecture to graph classification, which proves to be an efficient tool for improving base graph neural network architectures. Multiple improvements to baseline models are also proposed and experimentally verified, which constitutes an important contribution to the field of fair model comparison.
Feature selection is of great importance in Machine Learning, where it can be used to reduce the dimensionality of classification, ranking and prediction problems. The removal of redundant and noisy features can improve both the accuracy and scalability of the trained models. However, feature selection is a computationally expensive task with a solution space that grows combinatorically. In this work, we consider in particular a quadratic feature selection problem that can be tackled with the Quantum Approximate Optimization Algorithm (QAOA), already employed in combinatorial optimization. First we represent the feature selection problem with the QUBO formulation, which is then mapped to an Ising spin Hamiltonian. Then we apply QAOA with the goal of finding the ground state of this Hamiltonian, which corresponds to the optimal selection of features. In our experiments, we consider seven different real-world datasets with dimensionality up to 21 and run QAOA on both a quantum simulator and, for small datasets, the 7-qubit IBM (ibm-perth) quantum computer. We use the set of selected features to train a classification model and evaluate its accuracy. Our analysis shows that it is possible to tackle the feature selection problem with QAOA and that currently available quantum devices can be used effectively. Future studies could test a wider range of classification models as well as improve the effectiveness of QAOA by exploring better performing optimizers for its classical step.
Cross entropy loss has served as the main objective function for classification-based tasks. Widely deployed for learning neural network classifiers, it shows both effectiveness and a probabilistic interpretation. Recently, after the success of self supervised contrastive representation learning methods, supervised contrastive methods have been proposed to learn representations and have shown superior and more robust performance, compared to solely training with cross entropy loss. However, cross entropy loss is still needed to train the final classification layer. In this work, we investigate the possibility of learning both the representation and the classifier using one objective function that combines the robustness of contrastive learning and the probabilistic interpretation of cross entropy loss. First, we revisit a previously proposed contrastive-based objective function that approximates cross entropy loss and present a simple extension to learn the classifier jointly. Second, we propose a new version of the supervised contrastive training that learns jointly the parameters of the classifier and the backbone of the network. We empirically show that our proposed objective functions show a significant improvement over the standard cross entropy loss with more training stability and robustness in various challenging settings.
Recently, massive architectures based on Convolutional Neural Network (CNN) and self-attention mechanisms have become necessary for audio classification. While these techniques are state-of-the-art, these works' effectiveness can only be guaranteed with huge computational costs and parameters, large amounts of data augmentation, transfer from large datasets and some other tricks. By utilizing the lightweight nature of audio, we propose an efficient network structure called Paired Inverse Pyramid Structure (PIP) and a network called Paired Inverse Pyramid Structure MLP Network (PIPMN). The PIPMN reaches 96\% of Environmental Sound Classification (ESC) accuracy on the UrbanSound8K dataset and 93.2\% of Music Genre Classification (MGC) on the GTAZN dataset, with only 1 million parameters. Both of the results are achieved without data augmentation or model transfer. Public code is available at: https://github.com/JNAIC/PIPMN
The motor imagery (MI) classification has been a prominent research topic in brain-computer interfaces based on electroencephalography (EEG). Over the past few decades, the performance of MI-EEG classifiers has seen gradual enhancement. In this study, we amplify the geometric deep learning-based MI-EEG classifiers from the perspective of time-frequency analysis, introducing a new architecture called Graph-CSPNet. We refer to this category of classifiers as Geometric Classifiers, highlighting their foundation in differential geometry stemming from EEG spatial covariance matrices. Graph-CSPNet utilizes novel manifold-valued graph convolutional techniques to capture the EEG features in the time-frequency domain, offering heightened flexibility in signal segmentation for capturing localized fluctuations. To evaluate the effectiveness of Graph-CSPNet, we employ five commonly-used publicly available MI-EEG datasets, achieving near-optimal classification accuracies in nine out of eleven scenarios. The Python repository can be found at https://github.com/GeometricBCI/Tensor-CSPNet-and-Graph-CSPNet.
Let $\mathcal{D}$ be a dataset of smooth 3D-surfaces, partitioned into disjoint classes $\mathit{CL}_j$, $j= 1, \ldots, k$. We show how optimized diffeomorphic registration applied to large numbers of pairs $S,S' \in \mathcal{D}$ can provide descriptive feature vectors to implement automatic classification on $\mathcal{D}$, and generate classifiers invariant by rigid motions in $\mathbb{R}^3$. To enhance accuracy of automatic classification, we enrich the smallest classes $\mathit{CL}_j$ by diffeomorphic interpolation of smooth surfaces between pairs $S,S' \in \mathit{CL}_j$. We also implement small random perturbations of surfaces $S\in \mathit{CL}_j$ by random flows of smooth diffeomorphisms $F_t:\mathbb{R}^3 \to \mathbb{R}^3$. Finally, we test our automatic classification methods on a cardiology data base of discretized mitral valve surfaces.
We formulate the problem of performing optimal data compression under the constraints that compressed data can be used for accurate classification in machine learning. We show that this translates to a problem of minimizing the mutual information between data and its compressed version under the constraint on error probability of classification is small when using the compressed data for machine learning. We then provide analytical and computational methods to characterize the optimal trade-off between data compression and classification error probability. First, we provide an analytical characterization for the optimal compression strategy for data with binary labels. Second, for data with multiple labels, we formulate a set of convex optimization problems to characterize the optimal tradeoff, from which the optimal trade-off between the classification error and compression efficiency can be obtained by numerically solving the formulated optimization problems. We further show the improvements of our formulations over the information-bottleneck methods in classification performance.
Despite their remarkable performance, deep neural networks remain unadopted in clinical practice, which is considered to be partially due to their lack in explainability. In this work, we apply attribution methods to a pre-trained deep neural network (DNN) for 12-lead electrocardiography classification to open this "black box" and understand the relationship between model prediction and learned features. We classify data from a public data set and the attribution methods assign a "relevance score" to each sample of the classified signals. This allows analyzing what the network learned during training, for which we propose quantitative methods: average relevance scores over a) classes, b) leads, and c) average beats. The analyses of relevance scores for atrial fibrillation (AF) and left bundle branch block (LBBB) compared to healthy controls show that their mean values a) increase with higher classification probability and correspond to false classifications when around zero, and b) correspond to clinical recommendations regarding which lead to consider. Furthermore, c) visible P-waves and concordant T-waves result in clearly negative relevance scores in AF and LBBB classification, respectively. In summary, our analysis suggests that the DNN learned features similar to cardiology textbook knowledge.
To mitigate the bias exhibited by machine learning models, fairness criteria can be integrated into the training process to ensure fair treatment across all demographics, but it often comes at the expense of model performance. Understanding such tradeoffs, therefore, underlies the design of fair algorithms. To this end, this paper provides a complete characterization of the inherent tradeoff of demographic parity on classification problems, under the most general multi-group, multi-class, and noisy setting. Specifically, we show that the minimum error rate achievable by randomized and attribute-aware fair classifiers is given by the optimal value of a Wasserstein-barycenter problem. On the practical side, our findings lead to a simple post-processing algorithm that derives fair classifiers from score functions, which yields the optimal fair classifier when the score is Bayes optimal. We provide suboptimality analysis and sample complexity for our algorithm, and demonstrate its effectiveness on benchmark datasets.