Search

Membership inference attacks (MIAs) determine whether certain data instances were used to train a model by exploiting the differences in how the model responds to seen versus unseen instances. This capability makes MIAs important in assessing privacy leakage within modern generative AI systems. However, this paper reveals an oversight in existing MIAs against \emph{distilled generative models}: attackers can no longer detect a teacher model's training instances individually when targeting the distilled student model, as the student learns from the teacher-generated data rather than its original member data, preventing direct instance-level memorization. Nevertheless, we find that student-generated samples exhibit a significantly stronger distributional alignment with teacher's member data than non-member data. This leads us to posit that MIAs \emph{on distilled generative models should shift from instance-level to distribution-level statistics}. We thereby introduce a \emph{set-based} MIA framework that measures \emph{relative} distributional discrepancies between student-generated data\emph{sets} and potential member/non-member data\emph{sets}, Empirically, distributional statistics reliably distinguish a teacher's member data from non-member data through the distilled model. Finally, we discuss scenarios in which our setup faces limitations.

Concept drift, characterized by unpredictable changes in data distribution over time, poses significant challenges to machine learning models in streaming data scenarios. Although error rate-based concept drift detectors are widely used, they often fail to identify drift in the early stages when the data distribution changes but error rates remain constant. This paper introduces the Prediction Uncertainty Index (PU-index), derived from the prediction uncertainty of the classifier, as a superior alternative to the error rate for drift detection. Our theoretical analysis demonstrates that: (1) The PU-index can detect drift even when error rates remain stable. (2) Any change in the error rate will lead to a corresponding change in the PU-index. These properties make the PU-index a more sensitive and robust indicator for drift detection compared to existing methods. We also propose a PU-index-based Drift Detector (PUDD) that employs a novel Adaptive PU-index Bucketing algorithm for detecting drift. Empirical evaluations on both synthetic and real-world datasets demonstrate PUDD's efficacy in detecting drift in structured and image data., Comment: Accepted by AAAI-2025

Cross-Domain Recommendation (CDR) is a promising paradigm inspired by transfer learning to solve the cold-start problem in recommender systems. Existing state-of-the-art CDR methods train an explicit mapping function to transfer the cold-start users from a data-rich source domain to a target domain. However, a limitation of these methods is that the mapping function is trained on overlapping users across domains, while only a small number of overlapping users are available for training. By visualizing the loss landscape of the existing CDR model, we find that training on a small number of overlapping users causes the model to converge to sharp minima, leading to poor generalization. Based on this observation, we leverage loss-geometry-based machine learning approach and propose a novel CDR method called Sharpness-Aware CDR (SCDR). Our proposed method simultaneously optimizes recommendation loss and loss sharpness, leading to better generalization with theoretical guarantees. Empirical studies on real-world datasets demonstrate that SCDR significantly outperforms the other CDR models for cold-start recommendation tasks, while concurrently enhancing the model's robustness to adversarial attacks.

Deep reinforcement learning is used in various domains, but usually under the assumption that the environment has stationary conditions like transitions and state distributions. When this assumption is not met, performance suffers. For this reason, tracking continuous environmental changes and adapting to unpredictable conditions is challenging yet crucial because it ensures that systems remain reliable and flexible in practical scenarios. Our research introduces Behavior-Aware Detection and Adaptation (BADA), an innovative framework that merges environmental change detection with behavior adaptation. The key inspiration behind our method is that policies exhibit different global behaviors in changing environments. Specifically, environmental changes are identified by analyzing variations between behaviors using Wasserstein distances without manually set thresholds. The model adapts to the new environment through behavior regularization based on the extent of changes. The results of a series of experiments demonstrate better performance relative to several current algorithms. This research also indicates significant potential for tackling this long-standing challenge., Comment: Accepted by IJCAI 2024

Uncertain changes in data streams present challenges for machine learning models to dynamically adapt and uphold performance in real-time. Particularly, classification boundary change, also known as real concept drift, is the major cause of classification performance deterioration. However, accurately detecting real concept drift remains challenging because the theoretical foundations of existing drift detection methods - two-sample distribution tests and monitoring classification error rate, both suffer from inherent limitations such as the inability to distinguish virtual drift (changes not affecting the classification boundary, will introduce unnecessary model maintenance), limited statistical power, or high computational cost. Furthermore, no existing detection method can provide information on the trend of the drift, which could be invaluable for model maintenance. This work presents a novel real concept drift detection method based on Neighbor-Searching Discrepancy, a new statistic that measures the classification boundary difference between two samples. The proposed method is able to detect real concept drift with high accuracy while ignoring virtual drift. It can also indicate the direction of the classification boundary change by identifying the invasion or retreat of a certain class, which is also an indicator of separability change between classes. A comprehensive evaluation of 11 experiments is conducted, including empirical verification of the proposed theory using artificial datasets, and experimental comparisons with commonly used drift handling methods on real-world datasets. The results show that the proposed theory is robust against a range of distributions and dimensions, and the drift detection method outperforms state-of-the-art alternative methods.

Column generation (CG) is a well-established method for solving large-scale linear programs. It involves iteratively optimizing a subproblem containing a subset of columns and using its dual solution to generate new columns with negative reduced costs. This process continues until the dual values converge to the optimal dual solution to the original problem. A natural phenomenon in CG is the heavy oscillation of the dual values during iterations, which can lead to a substantial slowdown in the convergence rate. Stabilization techniques are devised to accelerate the convergence of dual values by using information beyond the state of the current subproblem. However, there remains a significant gap in obtaining more accurate dual values at an earlier stage. To further narrow this gap, this paper introduces a novel approach consisting of 1) a machine learning approach for accurate prediction of optimal dual solutions and 2) an adaptive stabilization technique that effectively capitalizes on accurate predictions. On the graph coloring problem, we show that our method achieves a significantly improved convergence rate compared to traditional methods., Comment: Accepted by ICML'24

Shifting the focus from principles to practical implementation, responsible artificial intelligence (AI) has garnered considerable attention across academia, industry, and society at large. Despite being in its nascent stages, this emerging field grapples with nebulous concepts and intricate knowledge frameworks. By analyzing three prevailing concepts - explainable AI, trustworthy AI, and ethical AI, this study defined responsible AI and identified its core principles. Methodologically, this study successfully demonstrated the implementation of leveraging AI's capabilities into bibliometrics for enhanced knowledge discovery and the cross-validation of experimentally examined models with domain insights. Empirically, this study investigated 17,799 research articles contributed by the AI community since 2015. This involves recognizing key technological players and their relationships, unveiling the topical landscape and hierarchy of responsible AI, charting its evolution, and elucidating the interplay between the responsibility principles and primary AI techniques. An analysis of a core cohort comprising 380 articles from multiple disciplines captures the most recent advancements in responsible AI. As one of the pioneering bibliometric studies dedicated to exploring responsible AI, this study will provide comprehensive macro-level insights, enhancing the understanding of responsible AI while furnishing valuable knowledge support for AI regulation and governance initiatives., Comment: 14 pages, 9 figures

Multistream classification poses significant challenges due to the necessity for rapid adaptation in dynamic streaming processes with concept drift. Despite the growing research outcomes in this area, there has been a notable oversight regarding the temporal dynamic relationships between these streams, leading to the issue of negative transfer arising from irrelevant data. In this paper, we propose a novel Online Boosting Adaptive Learning (OBAL) method that effectively addresses this limitation by adaptively learning the dynamic correlation among different streams. Specifically, OBAL operates in a dual-phase mechanism, in the first of which we design an Adaptive COvariate Shift Adaptation (AdaCOSA) algorithm to construct an initialized ensemble model using archived data from various source streams, thus mitigating the covariate shift while learning the dynamic correlations via an adaptive re-weighting strategy. During the online process, we employ a Gaussian Mixture Model-based weighting mechanism, which is seamlessly integrated with the acquired correlations via AdaCOSA to effectively handle asynchronous drift. This approach significantly improves the predictive performance and stability of the target stream. We conduct comprehensive experiments on several synthetic and real-world data streams, encompassing various drifting scenarios and types. The results clearly demonstrate that OBAL achieves remarkable advancements in addressing multistream classification problems by effectively leveraging positive knowledge derived from multiple sources., Comment: AAAI 2024

Out-of-distribution (OOD) detection is crucial to modern deep learning applications by identifying and alerting about the OOD samples that should not be tested or used for making predictions. Current OOD detection methods have made significant progress when in-distribution (ID) and OOD samples are drawn from static distributions. However, this can be unrealistic when applied to real-world systems which often undergo continuous variations and shifts in ID and OOD distributions over time. Therefore, for an effective application in real-world systems, the development of OOD detection methods that can adapt to these dynamic and evolving distributions is essential. In this paper, we propose a novel and more realistic setting called continuously adaptive out-of-distribution (CAOOD) detection which targets on developing an OOD detection model that enables dynamic and quick adaptation to a new arriving distribution, with insufficient ID samples during deployment time. To address CAOOD, we develop meta OOD learning (MOL) by designing a learning-to-adapt diagram such that a good initialized OOD detection model is learned during the training process. In the testing process, MOL ensures OOD detection performance over shifting distributions by quickly adapting to new distributions with a few adaptations. Extensive experiments on several OOD benchmarks endorse the effectiveness of our method in preserving both ID classification accuracy and OOD detection performance on continuously shifting distributions., Comment: Accepted by ICCV 2023

As a powerful Bayesian non-parameterized algorithm, the Gaussian process (GP) has performed a significant role in Bayesian optimization and signal processing. GPs have also advanced online decision-making systems because their posterior distribution has a closed-form solution. However, its training and inference process requires all historic data to be stored and the GP model to be trained from scratch. For those reasons, several online GP algorithms, such as O-SGPR and O-SVGP, have been specifically designed for streaming settings. In this paper, we present a new theoretical framework for online GPs based on the online probably approximately correct (PAC) Bayes theory. The framework offers both a guarantee of generalized performance and good accuracy. Instead of minimizing the marginal likelihood, our algorithm optimizes both the empirical risk function and a regularization item, which is in proportion to the divergence between the prior distribution and posterior distribution of parameters. In addition to its theoretical appeal, the algorithm performs well empirically on several regression datasets. Compared to other online GP algorithms, ours yields a generalization guarantee and very competitive accuracy.

The theoretical analysis of multi-class classification has proved that the existing multi-class classification methods can train a classifier with high classification accuracy on the test set, when the instances are precise in the training and test sets with same distribution and enough instances can be collected in the training set. However, one limitation with multi-class classification has not been solved: how to improve the classification accuracy of multi-class classification problems when only imprecise observations are available. Hence, in this paper, we propose a novel framework to address a new realistic problem called multi-class classification with imprecise observations (MCIMO), where we need to train a classifier with fuzzy-feature observations. Firstly, we give the theoretical analysis of the MCIMO problem based on fuzzy Rademacher complexity. Then, two practical algorithms based on support vector machine and neural networks are constructed to solve the proposed new problem. Experiments on both synthetic and real-world datasets verify the rationality of our theoretical analysis and the efficacy of the proposed algorithms., Comment: This article has been accepted by IEEE Transactions on Cybernetics on June 4, 2022

Transfer learning where the behavior of extracting transferable knowledge from the source domain(s) and reusing this knowledge to target domain has become a research area of great interest in the field of artificial intelligence. Probabilistic graphical models (PGMs) have been recognized as a powerful tool for modeling complex systems with many advantages, e.g., the ability to handle uncertainty and possessing good interpretability. Considering the success of these two aforementioned research areas, it seems natural to apply PGMs to transfer learning. However, although there are already some excellent PGMs specific to transfer learning in the literature, the potential of PGMs for this problem is still grossly underestimated. This paper aims to boost the development of PGMs for transfer learning by 1) examining the pilot studies on PGMs specific to transfer learning, i.e., analyzing and summarizing the existing mechanisms particularly designed for knowledge transfer; 2) discussing examples of real-world transfer problems where existing PGMs have been successfully applied; and 3) exploring several potential research directions on transfer learning using PGM.

Causal effect estimation for dynamic treatment regimes (DTRs) contributes to sequential decision making. However, censoring and time-dependent confounding under DTRs are challenging as the amount of observational data declines over time due to a reducing sample size but the feature dimension increases over time. Long-term follow-up compounds these challenges. Another challenge is the highly complex relationships between confounders, treatments, and outcomes, which causes the traditional and commonly used linear methods to fail. We combine outcome regression models with treatment models for high dimensional features using uncensored subjects that are small in sample size and we fit deep Bayesian models for outcome regression models to reveal the complex relationships between confounders, treatments, and outcomes. Also, the developed deep Bayesian models can model uncertainty and output the prediction variance which is essential for the safety-aware applications, such as self-driving cars and medical treatment design. The experimental results on medical simulations of HIV treatment show the ability of the proposed method to obtain stable and accurate dynamic causal effect estimation from observational data, especially with long-term follow-up. Our technique provides practical guidance for sequential decision making, and policy-making.

Traditional supervised learning aims to train a classifier in the closed-set world, where training and test samples share the same label space. In this paper, we target a more challenging and realistic setting: open-set learning (OSL), where there exist test samples from the classes that are unseen during training. Although researchers have designed many methods from the algorithmic perspectives, there are few methods that provide generalization guarantees on their ability to achieve consistent performance on different training samples drawn from the same distribution. Motivated by the transfer learning and probably approximate correct (PAC) theory, we make a bold attempt to study OSL by proving its generalization error-given training samples with size n, the estimation error will get close to order O_p(1/\sqrt{n}). This is the first study to provide a generalization bound for OSL, which we do by theoretically investigating the risk of the target classifier on unknown classes. According to our theory, a novel algorithm, called auxiliary open-set risk (AOSR) is proposed to address the OSL problem. Experiments verify the efficacy of AOSR. The code is available at github.com/Anjin-Liu/Openset_Learning_AOSR., Comment: Open-set Learning, Open-set Recognition, Machine Learning Theory

Many methods have been proposed to detect concept drift, i.e., the change in the distribution of streaming data, due to concept drift causes a decrease in the prediction accuracy of algorithms. However, the most of current detection methods are based on the assessment of the degree of change in the data distribution, cannot identify the type of concept drift. In this paper, we propose Active Drift Detection with Meta learning (Meta-ADD), a novel framework that learns to classify concept drift by tracking the changed pattern of error rates. Specifically, in the training phase, we extract meta-features based on the error rates of various concept drift, after which a meta-detector is developed via a prototypical neural network by representing various concept drift classes as corresponding prototypes. In the detection phase, the learned meta-detector is fine-tuned to adapt to the corresponding data stream via stream-based active learning. Hence, Meta-ADD uses machine learning to learn to detect concept drifts and identify their types automatically, which can directly support drift understand. The experiment results verify the effectiveness of Meta-ADD.

By leveraging experience from previous tasks, meta-learning algorithms can achieve effective fast adaptation ability when encountering new tasks. However it is unclear how the generalization property applies to new tasks. Probably approximately correct (PAC) Bayes bound theory provides a theoretical framework to analyze the generalization performance for meta-learning. We derive three novel generalisation error bounds for meta-learning based on PAC-Bayes relative entropy bound. Furthermore, using the empirical risk minimization (ERM) method, a PAC-Bayes bound for meta-learning with data-dependent prior is developed. Experiments illustrate that the proposed three PAC-Bayes bounds for meta-learning guarantee a competitive generalization performance guarantee, and the extended PAC-Bayes bound with data-dependent prior can achieve rapid convergence ability.

Unsupervised domain adaptation (UDA) aims to train a target classifier with labeled samples from the source domain and unlabeled samples from the target domain. Classical UDA learning bounds show that target risk is upper bounded by three terms: source risk, distribution discrepancy, and combined risk. Based on the assumption that the combined risk is a small fixed value, methods based on this bound train a target classifier by only minimizing estimators of the source risk and the distribution discrepancy. However, the combined risk may increase when minimizing both estimators, which makes the target risk uncontrollable. Hence the target classifier cannot achieve ideal performance if we fail to control the combined risk. To control the combined risk, the key challenge takes root in the unavailability of the labeled samples in the target domain. To address this key challenge, we propose a method named E-MixNet. E-MixNet employs enhanced mixup, a generic vicinal distribution, on the labeled source samples and pseudo-labeled target samples to calculate a proxy of the combined risk. Experiments show that the proxy can effectively curb the increase of the combined risk when minimizing the source risk and distribution discrepancy. Furthermore, we show that if the proxy of the combined risk is added into loss functions of four representative UDA methods, their performance is also improved., Comment: 9 pages, 3 figures, Accepted by Association for the Advancement of Artificial Intelligence 2021 (AAAI 2021)

In data streams, the data distribution of arriving observations at different time points may change - a phenomenon called concept drift. While detecting concept drift is a relatively mature area of study, solutions to the uncertainty introduced by observations with missing values have only been studied in isolation. No one has yet explored whether or how these solutions might impact drift detection performance. We, however, believe that data imputation methods may actually increase uncertainty in the data rather than reducing it. We also conjecture that imputation can introduce bias into the process of estimating distribution changes during drift detection, which can make it more difficult to train a learning model. Our idea is to focus on estimating the distance between observations rather than estimating the missing values, and to define membership functions that allocate observations to histogram bins according to the estimation errors. Our solution comprises a novel masked distance learning (MDL) algorithm to reduce the cumulative errors caused by iteratively estimating each missing value in an observation and a fuzzy-weighted frequency (FWF) method for identifying discrepancies in the data distribution. The concept drift detection algorithm proposed in this paper is a singular and unified algorithm that can handle missing values, but not an imputation algorithm combined with a concept drift detection algorithm. Experiments on both synthetic and real-world data sets demonstrate the advantages of this method and show its robustness in detecting drift in data with missing values. These findings reveal that missing values exert a profound impact on concept drift detection, but using fuzzy set theory to model observations can produce more reliable results than imputation., Comment: Accepted by IEEE Transactions on Fuzzy Systems

In unsupervised domain adaptation (UDA), a classifier for the target domain is trained with massive true-label data from the source domain and unlabeled data from the target domain. However, collecting fully-true-label data in the source domain is high-cost and sometimes impossible. Compared to the true labels, a complementary label specifies a class that a pattern does not belong to, hence collecting complementary labels would be less laborious than collecting true labels. Thus, in this paper, we propose a novel setting that the source domain is composed of complementary-label data, and a theoretical bound for it is first proved. We consider two cases of this setting, one is that the source domain only contains complementary-label data (completely complementary unsupervised domain adaptation, CC-UDA), and the other is that the source domain has plenty of complementary-label data and a small amount of true-label data (partly complementary unsupervised domain adaptation, PC-UDA). To this end, a complementary label adversarial network} (CLARINET) is proposed to solve CC-UDA and PC-UDA problems. CLARINET maintains two deep networks simultaneously, where one focuses on classifying complementary-label source data and the other takes care of source-to-target distributional adaptation. Experiments show that CLARINET significantly outperforms a series of competent baselines on handwritten-digits-recognition and objects-recognition tasks., Comment: arXiv admin note: text overlap with arXiv:2007.14612

In unsupervised domain adaptation (UDA), classifiers for the target domain are trained with massive true-label data from the source domain and unlabeled data from the target domain. However, it may be difficult to collect fully-true-label data in a source domain given a limited budget. To mitigate this problem, we consider a novel problem setting where the classifier for the target domain has to be trained with complementary-label data from the source domain and unlabeled data from the target domain named budget-friendly UDA (BFUDA). The key benefit is that it is much less costly to collect complementary-label source data (required by BFUDA) than collecting the true-label source data (required by ordinary UDA). To this end, the complementary label adversarial network (CLARINET) is proposed to solve the BFUDA problem. CLARINET maintains two deep networks simultaneously, where one focuses on classifying complementary-label source data and the other takes care of the source-to-target distributional adaptation. Experiments show that CLARINET significantly outperforms a series of competent baselines., Comment: This paper has been accepted by IJCAI-PRICAI 2020. Yiyang Zhang, Feng Liu and Zhen Fang equally contribute to this paper

In the unsupervised open set domain adaptation (UOSDA), the target domain contains unknown classes that are not observed in the source domain. Researchers in this area aim to train a classifier to accurately: 1) recognize unknown target data (data with unknown classes) and, 2) classify other target data. To achieve this aim, a previous study has proven an upper bound of the target-domain risk, and the open set difference, as an important term in the upper bound, is used to measure the risk on unknown target data. By minimizing the upper bound, a shallow classifier can be trained to achieve the aim. However, if the classifier is very flexible (e.g., deep neural networks (DNNs)), the open set difference will converge to a negative value when minimizing the upper bound, which causes an issue where most target data are recognized as unknown data. To address this issue, we propose a new upper bound of target-domain risk for UOSDA, which includes four terms: source-domain risk, $\epsilon$-open set difference ($\Delta_\epsilon$), a distributional discrepancy between domains, and a constant. Compared to the open set difference, $\Delta_\epsilon$ is more robust against the issue when it is being minimized, and thus we are able to use very flexible classifiers (i.e., DNNs). Then, we propose a new principle-guided deep UOSDA method that trains DNNs via minimizing the new upper bound. Specifically, source-domain risk and $\Delta_\epsilon$ are minimized by gradient descent, and the distributional discrepancy is minimized via a novel open-set conditional adversarial training strategy. Finally, compared to existing shallow and deep UOSDA methods, our method shows the state-of-the-art performance on several benchmark datasets, including digit recognition (MNIST, SVHN, USPS), object recognition (Office-31, Office-Home), and face recognition (PIE).

Data stream poses additional challenges to statistical classification tasks because distributions of the training and target samples may differ as time passes. Such distribution change in streaming data is called concept drift. Numerous histogram-based distribution change detection methods have been proposed to detect drift. Most histograms are developed on grid-based or tree-based space partitioning algorithms which makes the space partitions arbitrary, unexplainable, and may cause drift blind-spots. There is a need to improve the drift detection accuracy for histogram-based methods with the unsupervised setting. To address this problem, we propose a cluster-based histogram, called equal intensity k-means space partitioning (EI-kMeans). In addition, a heuristic method to improve the sensitivity of drift detection is introduced. The fundamental idea of improving the sensitivity is to minimize the risk of creating partitions in distribution offset regions. Pearson's chi-square test is used as the statistical hypothesis test so that the test statistics remain independent of the sample distribution. The number of bins and their shapes, which strongly influence the ability to detect drift, are determined dynamically from the sample based on an asymptotic constraint in the chi-square test. Accordingly, three algorithms are developed to implement concept drift detection, including a greedy centroids initialization algorithm, a cluster amplify-shrink algorithm, and a drift detection algorithm. For drift adaptation, we recommend retraining the learner if a drift is detected. The results of experiments on synthetic and real-world datasets demonstrate the advantages of EI-kMeans and show its efficacy in detecting concept drift.

Concept drift refers to changes in the distribution of underlying data and is an inherent property of evolving data streams. Ensemble learning, with dynamic classifiers, has proved to be an efficient method of handling concept drift. However, the best way to create and maintain ensemble diversity with evolving streams is still a challenging problem. In contrast to estimating diversity via inputs, outputs, or classifier parameters, we propose a diversity measurement based on whether the ensemble members agree on the probability of a regional distribution change. In our method, estimations over regional distribution changes are used as instance weights. Constructing different region sets through different schemes will lead to different drift estimation results, thereby creating diversity. The classifiers that disagree the most are selected to maximize diversity. Accordingly, an instance-based ensemble learning algorithm, called the diverse instance weighting ensemble (DiwE), is developed to address concept drift for data stream classification problems. Evaluations of various synthetic and real-world data stream benchmarks show the effectiveness and advantages of the proposed algorithm., Comment: in IEEE Transactions on Neural Networks and Learning Systems, 2020

Concept drift describes unforeseeable changes in the underlying distribution of streaming data over time. Concept drift research involves the development of methodologies and techniques for drift detection, understanding and adaptation. Data analysis has revealed that machine learning in a concept drift environment will result in poor learning results if the drift is not addressed. To help researchers identify which research topics are significant and how to apply related techniques in data analysis tasks, it is necessary that a high quality, instructive review of current research developments and trends in the concept drift field is conducted. In addition, due to the rapid development of concept drift in recent years, the methodologies of learning under concept drift have become noticeably systematic, unveiling a framework which has not been mentioned in literature. This paper reviews over 130 high quality publications in concept drift related research areas, analyzes up-to-date developments in methodologies and techniques, and establishes a framework of learning under concept drift including three main components: concept drift detection, concept drift understanding, and concept drift adaptation. This paper lists and discusses 10 popular synthetic datasets and 14 publicly available benchmark datasets used for evaluating the performance of learning algorithms aiming at handling concept drift. Also, concept drift related research directions are covered and discussed. By providing state-of-the-art knowledge, this survey will directly support researchers in their understanding of research developments in the field of learning under concept drift.

We propose a class of kernel-based two-sample tests, which aim to determine whether two sets of samples are drawn from the same distribution. Our tests are constructed from kernels parameterized by deep neural nets, trained to maximize test power. These tests adapt to variations in distribution smoothness and shape over space, and are especially suited to high dimensions and complex data. By contrast, the simpler kernels used in prior kernel testing work are spatially homogeneous, and adaptive only in lengthscale. We explain how this scheme includes popular classifier-based two-sample tests as a special case, but improves on them in general. We provide the first proof of consistency for the proposed adaptation method, which applies both to kernels on deep features and to simpler radial basis kernels or multiple kernel learning. In experiments, we establish the superior performance of our deep kernels in hypothesis testing on benchmark and real-world data. The code of our deep-kernel-based two sample tests is available at https://github.com/fengliu90/DK-for-TST.