Generative Adversarial Networks (GANs) is a powerful family of models that learn an underlying distribution to generate synthetic data. Many existing studies of GANs focus on improving the realness of the generated image data for visual applications, and few of them concern about improving the quality of the generated data for training other classifiers -- a task known as the model compatibility problem. As a consequence, existing GANs often prefer generating `easier' synthetic data that are far from the boundaries of the classifiers, and refrain from generating near-boundary data, which are known to play an important roles in training the classifiers. To improve GAN in terms of model compatibility, we propose Boundary-Calibration GANs (BCGANs), which leverage the boundary information from a set of pre-trained classifiers using the original data. In particular, we introduce an auxiliary Boundary-Calibration loss (BC-loss) into the generator of GAN to match the statistics between the posterior distributions of original data and generated data with respect to the boundaries of the pre-trained classifiers. The BC-loss is provably unbiased and can be easily coupled with different GAN variants to improve their model compatibility. Experimental results demonstrate that BCGANs not only generate realistic images like original GANs but also achieves superior model compatibility than the original GANs.
Conditional Generative Adversarial Networks (cGANs) are implicit generative models which allow to sample from class-conditional distributions. Existing cGANs are based on a wide range of different discriminator designs and training objectives. One popular design in earlier works is to include a classifier during training with the assumption that good classifiers can help eliminate samples generated with wrong classes. Nevertheless, including classifiers in cGANs often comes with a side effect of only generating easy-to-classify samples. Recently, some representative cGANs avoid the shortcoming and reach state-of-the-art performance without having classifiers. Somehow it remains unanswered whether the classifiers can be resurrected to design better cGANs. In this work, we demonstrate that classifiers can be properly leveraged to improve cGANs. We start by using the decomposition of the joint probability distribution to connect the goals of cGANs and classification as a unified framework. The framework, along with a classic energy model to parameterize distributions, justifies the use of classifiers for cGANs in a principled manner. It explains several popular cGAN variants, such as ACGAN, ProjGAN, and ContraGAN, as special cases with different levels of approximations, which provides a unified view and brings new insights to understanding cGANs. Experimental results demonstrate that the design inspired by the proposed framework outperforms state-of-the-art cGANs on multiple benchmark datasets, especially on the most challenging ImageNet. The code is available at https://github.com/sian-chen/PyTorch-ECGAN.
The goal of multi-label learning (MLL) is to associate a given instance with its relevant labels from a set of concepts. Previous works of MLL mainly focused on the setting where the concept set is assumed to be fixed, while many real-world applications require introducing new concepts into the set to meet new demands. One common need is to refine the original coarse concepts and split them into finer-grained ones, where the refinement process typically begins with limited labeled data for the finer-grained concepts. To address the need, we formalize the problem into a special weakly supervised MLL problem to not only learn the fine-grained concepts efficiently but also allow interactive queries to strategically collect more informative annotations to further improve the classifier. The key idea within our approach is to learn to assign pseudo-labels to the unlabeled entries, and in turn leverage the pseudo-labels to train the underlying classifier and to inform a better query strategy. Experimental results demonstrate that our pseudo-label approach is able to accurately recover the missing ground truth, boosting the prediction performance significantly over the baseline methods and facilitating a competitive active learning strategy.
Many recent developments on generative models for natural images have relied on heuristically-motivated metrics that can be easily gamed by memorizing a small sample from the true distribution or training a model directly to improve the metric. In this work, we critically evaluate the gameability of these metrics by designing and deploying a generative modeling competition. Our competition received over 11000 submitted models. The competitiveness between participants allowed us to investigate both intentional and unintentional memorization in generative modeling. To detect intentional memorization, we propose the ``Memorization-Informed Fr\'echet Inception Distance'' (MiFID) as a new memorization-aware metric and design benchmark procedures to ensure that winning submissions made genuine improvements in perceptual quality. Furthermore, we manually inspect the code for the 1000 top-performing models to understand and label different forms of memorization. Our analysis reveals that unintentional memorization is a serious and common issue in popular generative models. The generated images and our memorization labels of those models as well as code to compute MiFID are released to facilitate future studies on benchmarking generative models.
Precipitation nowcasting is an important task for weather forecasting. Many recent works aim to predict the high rainfall events more accurately with the help of deep learning techniques, but such events are relatively rare. The rarity is often addressed by formulations that re-weight the rare events. Somehow such a formulation carries a side effect of making "blurry" predictions in low rainfall regions and cannot convince meteorologists to trust its practical usability. We fix the trust issue by introducing a discriminator that encourages the prediction model to generate realistic rain-maps without sacrificing predictive accuracy. Furthermore, we extend the nowcasting time frame from one hour to three hours to further address the needs from meteorologists. The extension is based on consecutive attentions across different hours. We propose a new deep learning model for precipitation nowcasting that includes both the discrimination and attention techniques. The model is examined on a newly-built benchmark dataset that contains both radar data and actual rain data. The benchmark, which will be publicly released, not only establishes the superiority of the proposed model, but also is expected to encourage future research on precipitation nowcasting.
Active learning strives to reduce annotation costs by choosing the most critical examples to label. Typically, the active learning strategy is contingent on the classification model. For instance, uncertainty sampling depends on poorly calibrated model confidence scores. In the cold-start setting, active learning is impractical because of model instability and data scarcity. Fortunately, modern NLP provides an additional source of information: pre-trained language models. The pre-training loss can find examples that surprise the model and should be labeled for efficient fine-tuning. Therefore, we treat the language modeling loss as a proxy for classification uncertainty. With BERT, we develop a simple strategy based on the masked language modeling loss that minimizes labeling costs for text classification. Compared to other baselines, our approach reaches higher accuracy within less sampling iterations and computation time.
Gaze estimation involves predicting where the person is looking at, given either a single input image or a sequence of images. One challenging task, gaze estimation in the wild, concerns data collected in unconstrained environments with varying camera-person distances, like the Gaze360 dataset. The varying distances result in varying face sizes in the images, which makes it hard for current CNN backbones to estimate the gaze robustly. Inspired by our natural skill to identify the gaze by taking a focused look at the face area, we propose a novel architecture that similarly zooms in on the face area of the image at multiple scales to improve prediction accuracy. Another challenging task, 360-degree gaze estimation (also introduced by the Gaze360 dataset), consists of estimating not only the forward gazes, but also the backward ones. The backward gazes introduce discontinuity in the yaw angle values of the gaze, making the deep learning models affected by some huge loss around the discontinuous points. We propose to convert the angle values by sine-cosine transform to avoid the discontinuity and represent the physical meaning of the yaw angle better. We conduct ablation studies on both ideas, the novel architecture and the transform, to validate their effectiveness. The two ideas allow our proposed model to achieve state-of-the-art performance for both the Gaze360 dataset and the RT-Gene dataset when using single images. Furthermore, we extend the model to a sequential version that systematically zooms in on a given sequence of images. The sequential version again achieves state-of-the-art performance on the Gaze360 dataset, which further demonstrates the usefulness of our proposed ideas.
In weakly supervised learning, unbiased risk estimator(URE) is a powerful tool for training classifiers when training and test data are drawn from different distributions. Nevertheless, UREs lead to overfitting in many problem settings when the models are complex like deep networks. In this paper, we investigate reasons for such overfitting by studying a weakly supervised problem called learning with complementary labels. We argue the quality of gradient estimation matters more in risk minimization. Theoretically, we show that a URE gives an unbiased gradient estimator(UGE). Practically, however, UGEs may suffer from huge variance, which causes empirical gradients to be usually far away from true gradients during minimization. To this end, we propose a novel surrogate complementary loss(SCL) framework that trades zero bias with reduced variance and makes empirical gradients more aligned with true gradients in the direction. Thanks to this characteristic, SCL successfully mitigates the overfitting issue and improves URE-based methods.