The attention mechanism has been widely used in deep neural networks as a model component. By now, it has become a critical building block in many state-of-the-art natural language models. Despite its great success established empirically, the working mechanism of attention has not been investigated at a sufficient theoretical depth to date. In this paper, we set up a simple text classification task and study the dynamics of training a simple attention-based classification model using gradient descent. In this setting, we show that, for the discriminative words that the model should attend to, a persisting identity exists relating its embedding and the inner product of its key and the query. This allows us to prove that training must converge to attending to the discriminative words when the attention output is classified by a linear classifier. Experiments are performed, which validate our theoretical analysis and provide further insights.
We address the challenging problem of whole slide image (WSI) classification. WSIs have very high resolutions and usually lack localized annotations. WSI classification can be cast as a multiple instance learning (MIL) problem when only slide-level labels are available. We propose a MIL-based method for WSI classification and tumor detection that does not require localized annotations. Our method has three major components. First, we introduce a novel MIL aggregator that models the relations of the instances in a dual-stream architecture with trainable distance measurement. Second, since WSIs can produce large or unbalanced bags that hinder the training of MIL models, we propose to use self-supervised contrastive learning to extract good representations for MIL and alleviate the issue of prohibitive memory cost for large bags. Third, we adopt a pyramidal fusion mechanism for multiscale WSI features, and further improve the accuracy of classification and localization. Our model is evaluated on two representative WSI datasets. The classification accuracy of our model compares favorably to fully-supervised methods, with less than 2% accuracy gap across datasets. Our results also outperform all previous MIL-based methods. Additional benchmark results on standard MIL datasets further demonstrate the superior performance of our MIL aggregator on general MIL problems. GitHub repository: https://github.com/binli123/dsmil-wsi
Explanation methods facilitate the development of models that learn meaningful concepts and avoid exploiting spurious correlations. We illustrate a previously unrecognized limitation of the popular neural network explanation method Grad-CAM: as a side effect of the gradient averaging step, Grad-CAM sometimes highlights locations the model did not actually use. To solve this problem, we propose HiResCAM, a novel class-specific explanation method that is guaranteed to highlight only the locations the model used to make each prediction. We prove that HiResCAM is a generalization of CAM and explore the relationships between HiResCAM and other gradient-based explanation methods. Experiments on PASCAL VOC 2012, including crowd-sourced evaluations, illustrate that while HiResCAM's explanations faithfully reflect the model, Grad-CAM often expands the attention to create bigger and smoother visualizations. Overall, this work advances convolutional neural network explanation approaches and may aid in the development of trustworthy models for sensitive applications.
Whether what you see in Figure 1 is a "flamingo" or a "bird", is the question we ask in this paper. While fine-grained visual classification (FGVC) strives to arrive at the former, for the majority of us non-experts just "bird" would probably suffice. The real question is therefore -- how can we tailor for different fine-grained definitions under divergent levels of expertise. For that, we re-envisage the traditional setting of FGVC, from single-label classification, to that of top-down traversal of a pre-defined coarse-to-fine label hierarchy -- so that our answer becomes "bird"-->"Phoenicopteriformes"-->"Phoenicopteridae"-->"flamingo". To approach this new problem, we first conduct a comprehensive human study where we confirm that most participants prefer multi-granularity labels, regardless whether they consider themselves experts. We then discover the key intuition that: coarse-level label prediction exacerbates fine-grained feature learning, yet fine-level feature betters the learning of coarse-level classifier. This discovery enables us to design a very simple albeit surprisingly effective solution to our new problem, where we (i) leverage level-specific classification heads to disentangle coarse-level features with fine-grained ones, and (ii) allow finer-grained features to participate in coarser-grained label predictions, which in turn helps with better disentanglement. Experiments show that our method achieves superior performance in the new FGVC setting, and performs better than state-of-the-art on traditional single-label FGVC problem as well. Thanks to its simplicity, our method can be easily implemented on top of any existing FGVC frameworks and is parameter-free.
The goal of semi-supervised learning is to utilize the unlabeled, in-domain dataset U to improve models trained on the labeled dataset D. Under the context of large-scale language-model (LM) pretraining, how we can make the best use of U is poorly understood: is semi-supervised learning still beneficial with the presence of large-scale pretraining? should U be used for in-domain LM pretraining or pseudo-label generation? how should the pseudo-label based semi-supervised model be actually implemented? how different semi-supervised strategies affect performances regarding D of different sizes, U of different sizes, etc. In this paper, we conduct comprehensive studies on semi-supervised learning in the task of text classification under the context of large-scale LM pretraining. Our studies shed important lights on the behavior of semi-supervised learning methods: (1) with the presence of in-domain pretraining LM on U, open-domain LM pretraining is unnecessary; (2) both the in-domain pretraining strategy and the pseudo-label based strategy introduce significant performance boosts, with the former performing better with larger U, the latter performing better with smaller U, and the combination leading to the largest performance boost; (3) self-training (pretraining first on pseudo labels D' and then fine-tuning on D) yields better performances when D is small, while joint training on the combination of pseudo labels D' and the original dataset D yields better performances when D is large. Using semi-supervised learning strategies, we are able to achieve a performance of around 93.8% accuracy with only 50 training data points on the IMDB dataset, and a competitive performance of 96.6% with the full IMDB dataset. Our work marks an initial step in understanding the behavior of semi-supervised learning models under the context of large-scale pretraining.
We study out-of-distribution (OOD) prediction behavior of neural networks when they classify images from unseen classes or corrupted images. To probe the OOD behavior, we introduce a new measure, nearest category generalization (NCG), where we compute the fraction of OOD inputs that are classified with the same label as their nearest neighbor in the training set. Our motivation stems from understanding the prediction patterns of adversarially robust networks, since previous work has identified unexpected consequences of training to be robust to norm-bounded perturbations. We find that robust networks have consistently higher NCG accuracy than natural training, even when the OOD data is much farther away than the robustness radius. This implies that the local regularization of robust training has a significant impact on the network's decision regions. We replicate our findings using many datasets, comparing new and existing training methods. Overall, adversarially robust networks resemble a nearest neighbor classifier when it comes to OOD data. Code available at https://github.com/yangarbiter/nearest-category-generalization.
This paper presents the 'hyper-sinh', a variation of the m-arcsinh activation function suitable for Deep Learning (DL)-based algorithms for supervised learning, such as Convolutional Neural Networks (CNN). hyper-sinh, developed in the open source Python libraries TensorFlow and Keras, is thus described and validated as an accurate and reliable activation function for both shallow and deep neural networks. Improvements in accuracy and reliability in image and text classification tasks on five (N = 5) benchmark data sets available from Keras are discussed. Experimental results demonstrate the overall competitive classification performance of both shallow and deep neural networks, obtained via this novel function. This function is evaluated with respect to gold standard activation functions, demonstrating its overall competitive accuracy and reliability for both image and text classification.
Various evaluation measures have been developed for multi-label classification, including Hamming Loss (HL), Subset Accuracy (SA) and Ranking Loss (RL). However, there is a gap between empirical results and the existing theories: 1) an algorithm often empirically performs well on some measure(s) while poorly on others, while a formal theoretical analysis is lacking; and 2) in small label space cases, the algorithms optimizing HL often have comparable or even better performance on the SA measure than those optimizing SA directly, while existing theoretical results show that SA and HL are conflicting measures. This paper provides an attempt to fill up this gap by analyzing the learning guarantees of the corresponding learning algorithms on both SA and HL measures. We show that when a learning algorithm optimizes HL with its surrogate loss, it enjoys an error bound for the HL measure independent of $c$ (the number of labels), while the bound for the SA measure depends on at most $O(c)$. On the other hand, when directly optimizing SA with its surrogate loss, it has learning guarantees that depend on $O(\sqrt{c})$ for both HL and SA measures. This explains the observation that when the label space is not large, optimizing HL with its surrogate loss can have promising performance for SA. We further show that our techniques are applicable to analyze the learning guarantees of algorithms on other measures, such as RL. Finally, the theoretical analyses are supported by experimental results.
Weakly-supervised learning (WSL) has recently triggered substantial interest as it mitigates the lack of pixel-wise annotations. Given global image labels, WSL methods yield pixel-level predictions (segmentations), which enable to interpret class predictions. Despite their recent success, mostly with natural images, such methods can face important challenges when the foreground and background regions have similar visual cues, yielding high false-positive rates in segmentations, as is the case in challenging histology images. WSL training is commonly driven by standard classification losses, which implicitly maximize model confidence, and locate the discriminative regions linked to classification decisions. Therefore, they lack mechanisms for modeling explicitly non-discriminative regions and reducing false-positive rates. We propose novel regularization terms, which enable the model to seek both non-discriminative and discriminative regions, while discouraging unbalanced segmentations. We introduce high uncertainty as a criterion to localize non-discriminative regions that do not affect classifier decision, and describe it with original Kullback-Leibler (KL) divergence losses evaluating the deviation of posterior predictions from the uniform distribution. Our KL terms encourage high uncertainty of the model when the latter inputs the latent non-discriminative regions. Our loss integrates: (i) a cross-entropy seeking a foreground, where model confidence about class prediction is high; (ii) a KL regularizer seeking a background, where model uncertainty is high; and (iii) log-barrier terms discouraging unbalanced segmentations. Comprehensive experiments and ablation studies over the public GlaS colon cancer data and a Camelyon16 patch-based benchmark for breast cancer show substantial improvements over state-of-the-art WSL methods, and confirm the effect of our new regularizers.
The recently introduced locally orderless tensor network (LoTeNet) for supervised image classification uses matrix product state (MPS) operations on grids of transformed image patches. The resulting patch representations are combined back together into the image space and aggregated hierarchically using multiple MPS blocks per layer to obtain the final decision rules. In this work, we propose a non-patch based modification to LoTeNet that performs one MPS operation per layer, instead of several patch-level operations. The spatial information in the input images to MPS blocks at each layer is squeezed into the feature dimension, similar to LoTeNet, to maximise retained spatial correlation between pixels when images are flattened into 1D vectors. The proposed multi-layered tensor network (MLTN) is capable of learning linear decision boundaries in high dimensional spaces in a multi-layered setting, which results in a reduction in the computation cost compared to LoTeNet without any degradation in performance.
Attention maps are a popular way of explaining the decisions of convolutional networks for image classification. Typically, for each image of interest, a single attention map is produced, which assigns weights to pixels based on their importance to the classification. A single attention map, however, provides an incomplete understanding since there are often many other maps that explain a classification equally well. In this paper, we introduce structured attention graphs (SAGs), which compactly represent sets of attention maps for an image by capturing how different combinations of image regions impact a classifier's confidence. We propose an approach to compute SAGs and a visualization for SAGs so that deeper insight can be gained into a classifier's decisions. We conduct a user study comparing the use of SAGs to traditional attention maps for answering counterfactual questions about image classifications. Our results show that the users are more correct when answering comparative counterfactual questions based on SAGs compared to the baselines.
Robustness is of central importance in machine learning and has given rise to the fields of domain generalization and invariant learning, which are concerned with improving performance on a test distribution distinct from but related to the training distribution. In light of recent work suggesting an intimate connection between fairness and robustness, we investigate whether algorithms from robust ML can be used to improve the fairness of classifiers that are trained on biased data and tested on unbiased data. We apply Invariant Risk Minimization (IRM), a domain generalization algorithm that employs a causal discovery inspired method to find robust predictors, to the task of fairly predicting the toxicity of internet comments. We show that IRM achieves better out-of-distribution accuracy and fairness than Empirical Risk Minimization (ERM) methods, and analyze both the difficulties that arise when applying IRM in practice and the conditions under which IRM will likely be effective in this scenario. We hope that this work will inspire further studies of how robust machine learning methods relate to algorithmic fairness.
Previous studies on image classification have mainly focused on the performance of the networks, not on real-time operation or model compression. We propose a Gaussian Deep Recurrent visual Attention Model (GDRAM)- a reinforcement learning based lightweight deep neural network for large scale image classification that outperforms the conventional CNN (Convolutional Neural Network) which uses the entire image as input. Highly inspired by the biological visual recognition process, our model mimics the stochastic location of the retina with Gaussian distribution. We evaluate the model on Large cluttered MNIST, Large CIFAR-10 and Large CIFAR-100 datasets which are resized to 128 in both width and height.
It is challenging to visually detect heart disease from the electrocardiographic (ECG) signals. Implementing an automated ECG signal detection system can help diagnosis arrhythmia in order to improve the accuracy of diagnosis. In this paper, we proposed, implemented, and compared an automated system using two different frameworks of the combination of convolutional neural network (CNN) and long-short term memory (LSTM) for classifying normal sinus signals, atrial fibrillation, and other noisy signals. The dataset we used is from the MIT-BIT Arrhythmia Physionet. Our approach demonstrated that the cascade of two deep learning network has higher performance than the concatenation of them, achieving a weighted f1 score of 0.82. The experimental results have successfully validated that the cascade of CNN and LSTM can achieve satisfactory performance on discriminating ECG signals.
Deep learning techniques have provided significant improvements in hyperspectral image (HSI) classification. The current deep learning based HSI classifiers follow a patch-based learning framework by dividing the image into overlapping patches. As such, these methods are local learning methods, which have a high computational cost. In this paper, a fast patch-free global learning (FPGA) framework is proposed for HSI classification. In FPGA, an encoder-decoder based FCN is utilized to consider the global spatial information by processing the whole image, which results in fast inference. However, it is difficult to directly utilize the encoder-decoder based FCN for HSI classification as it always fails to converge due to the insufficiently diverse gradients caused by the limited training samples. To solve the divergence problem and maintain the abilities of FCN of fast inference and global spatial information mining, a global stochastic stratified sampling strategy is first proposed by transforming all the training samples into a stochastic sequence of stratified samples. This strategy can obtain diverse gradients to guarantee the convergence of the FCN in the FPGA framework. For a better design of FCN architecture, FreeNet, which is a fully end-to-end network for HSI classification, is proposed to maximize the exploitation of the global spatial information and boost the performance via a spectral attention based encoder and a lightweight decoder. A lateral connection module is also designed to connect the encoder and decoder, fusing the spatial details in the encoder and the semantic features in the decoder. The experimental results obtained using three public benchmark datasets suggest that the FPGA framework is superior to the patch-based framework in both speed and accuracy for HSI classification. Code has been made available at: https://github.com/Z-Zheng/FreeNet.
Loss functions is a crucial factor that affecting the detection precision in object detection task. In this paper, we optimize both two loss functions for classification and localization simultaneously. Firstly, by multiplying an IoU-based coefficient by the standard cross entropy loss in classification loss function, the correlation between localization and classification is established. Compared to the existing studies, in which the correlation is only applied to improve the localization accuracy for positive samples, this paper utilizes the correlation to obtain the really hard negative samples and aims to decrease the misclassified rate for negative samples. Besides, a novel localization loss named MIoU is proposed by incorporating a Mahalanobis distance between predicted box and target box, which eliminate the gradients inconsistency problem in the DIoU loss, further improving the localization accuracy. Finally, sufficient experiments for nighttime vehicle detection have been done on two datasets. Our results show than when train with the proposed loss functions, the detection performance can be outstandingly improved. The source code and trained models are available at https://github.com/therebellll/NegIoU-PosIoU-Miou.
Solving classification with graph methods has gained huge popularity in recent years. This is due to the fact that the data can be intuitively modeled with graphs to utilize high level features to aid in solving the classification problem. CULP which is short for Classification Using Link Prediction is a graph-based classifier. This classifier utilizes the graph representation of the data and transforms the problem to that of link prediction where we try to find the link between an unlabeled node and the proper class node for it. CULP proved to be highly accurate classifier and it has the power to predict the labels in near constant time. A variant of the classification problem is multi-label classification which tackles this problem for multi-label data where an instance can have multiple labels associated to it. In this work, we extend the CULP algorithm to address this problem. Our proposed extensions conveys the powers of CULP and its intuitive representation of the data in to the multi-label domain and in comparison to some of the cutting edge multi-label classifiers, yield competitive results.
Classification of polarimetric synthetic aperture radar (PolSAR) images is an active research area with a major role in environmental applications. The traditional Machine Learning (ML) methods proposed in this domain generally focus on utilizing highly discriminative features to improve the classification performance, but this task is complicated by the well-known "curse of dimensionality" phenomena. Other approaches based on deep Convolutional Neural Networks (CNNs) have certain limitations and drawbacks, such as high computational complexity, an unfeasibly large training set with ground-truth labels, and special hardware requirements. In this work, to address the limitations of traditional ML and deep CNN based methods, a novel and systematic classification framework is proposed for the classification of PolSAR images, based on a compact and adaptive implementation of CNNs using a sliding-window classification approach. The proposed approach has three advantages. First, there is no requirement for an extensive feature extraction process. Second, it is computationally efficient due to utilized compact configurations. In particular, the proposed compact and adaptive CNN model is designed to achieve the maximum classification accuracy with minimum training and computational complexity. This is of considerable importance considering the high costs involved in labelling in PolSAR classification. Finally, the proposed approach can perform classification using smaller window sizes than deep CNNs. Experimental evaluations have been performed over the most commonly-used four benchmark PolSAR images: AIRSAR L-Band and RADARSAT-2 C-Band data of San Francisco Bay and Flevoland areas. Accordingly, the best obtained overall accuracies range between 92.33 - 99.39% for these benchmark study sites.
Social coding platforms, such as GitHub, serve as laboratories for studying collaborative problem solving in open source software development; a key feature is their ability to support issue reporting which is used by teams to discuss tasks and ideas. Analyzing the dialogue between team members, as expressed in issue comments, can yield important insights about the performance of virtual teams. This paper presents a transfer learning approach for performing dialogue act classification on issue comments. Since no large labeled corpus of GitHub issue comments exists, employing transfer learning enables us to leverage standard dialogue act datasets in combination with our own GitHub comment dataset. We compare the performance of several word and sentence level encoding models including Global Vectors for Word Representations (GloVe), Universal Sentence Encoder (USE), and Bidirectional Encoder Representations from Transformers (BERT). Being able to map the issue comments to dialogue acts is a useful stepping stone towards understanding cognitive team processes.
Convolutional networks are at the center of best-in-class computer vision applications for a wide assortment of undertakings. Since 2014, a profound amount of work began to make better convolutional architectures, yielding generous additions in different benchmarks. Albeit expanded model size and computational cost will, in general, mean prompt quality increases for most undertakings but, the architectures now need to have some additional information to increase the performance. I show evidence that with the amalgamation of content-based image similarity and deep learning models, we can provide the flow of information which can be used in making clustered learning possible. The paper shows how training of sub-dataset clusters not only reduces the cost of computation but also increases the speed of evaluating and tuning a model on the given dataset.