The generalization of neural networks is a central challenge in machine learning, especially concerning the performance under distributions that differ from training ones. Current methods, mainly based on the data-driven paradigm such as data augmentation, adversarial training, and noise injection, may encounter limited generalization due to model non-smoothness. In this paper, we propose to investigate generalization from a Partial Differential Equation (PDE) perspective, aiming to enhance it directly through the underlying function of neural networks, rather than focusing on adjusting input data. Specifically, we first establish the connection between neural network generalization and the smoothness of the solution to a specific PDE, namely transport equation. Building upon this, we propose a general framework that introduces adaptive distributional diffusion into transport equation to enhance the smoothness of its solution, thereby improving generalization. In the context of neural networks, we put this theoretical framework into practice as PDE+ (PDE with Adaptive Distributional Diffusion) which diffuses each sample into a distribution covering semantically similar inputs. This enables better coverage of potentially unobserved distributions in training, thus improving generalization beyond merely data-driven methods. The effectiveness of PDE+ is validated through extensive experimental settings, demonstrating its superior performance compared to state-of-the-art methods. Our code is available at https://github.com/yuanyige/pde-add.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Due to the recent success of diffusion models, text-to-image generation is becoming increasingly popular and achieves a wide range of applications. Among them, text-to-image editing, or continuous text-to-image generation, attracts lots of attention and can potentially improve the quality of generated images. It's common to see that users may want to slightly edit the generated image by making minor modifications to their input textual descriptions for several rounds of diffusion inference. However, such an image editing process suffers from the low inference efficiency of many existing diffusion models even using GPU accelerators. To solve this problem, we introduce Fast Image Semantically Edit (FISEdit), a cached-enabled sparse diffusion model inference engine for efficient text-to-image editing. The key intuition behind our approach is to utilize the semantic mapping between the minor modifications on the input text and the affected regions on the output image. For each text editing step, FISEdit can 1) automatically identify the affected image regions and 2) utilize the cached unchanged regions' feature map to accelerate the inference process. For the former, we measure the differences between cached and ad hoc feature maps given the modified textual description, extract the region with significant differences, and capture the affected region by masks. For the latter, we develop an efficient sparse diffusion inference engine that only computes the feature maps for the affected region while reusing the cached statistics for the rest of the image. Finally, extensive empirical results show that FISEdit can be 3.4 times and 4.4 times faster than existing methods on NVIDIA TITAN RTX and A100 GPUs respectively, and even generates more satisfactory images.
Continual Semantic Segmentation (CSS) is an emerging trend, where catastrophic forgetting has been a perplexing problem. In this paper, we propose a Text-to-Image Knowledge Preservation (TIKP) framework to address this issue. TIKP applies Text-to-Image techniques to CSS by automatically generating prompts and content adaptation. It extracts associations between the labels of seen data and constructs text-level prompts based on these associations, which are preserved and maintained at each incremental step. During training, these prompts generate correlated images to mitigate the catastrophic forgetting. Particularly, as the generated images may have different distributions from the original data, TIKP transfers the knowledge by a content adaption loss, which determines the role played by the generated images in incremental training based on the similarity. In addition, for the classifier, we use the previous model from a different perspective: misclassifying new classes into old objects instead of the background. We propose a knowledge distillation loss based on wrong labels, enabling us to attribute varying weights to individual objects during the distillation process. Extensive experiments conducted in the same setting show that TIKP outperforms state-of-the-art methods by a large margin on benchmark datasets.
Semi-supervised learning (SSL) methods assume that labeled data, unlabeled data and test data are from the same distribution. Open-set semi-supervised learning (Open-set SSL) con- siders a more practical scenario, where unlabeled data and test data contain new categories (outliers) not observed in labeled data (inliers). Most previous works focused on out- lier detection via binary classifiers, which suffer from insufficient scalability and inability to distinguish different types of uncertainty. In this paper, we propose a novel framework, Adaptive Negative Evidential Deep Learning (ANEDL) to tackle these limitations. Concretely, we first introduce evidential deep learning (EDL) as an outlier detector to quantify different types of uncertainty, and design different uncertainty metrics for self-training and inference. Furthermore, we propose a novel adaptive negative optimization strategy, making EDL more tailored to the unlabeled dataset containing both inliers and outliers. As demonstrated empirically, our proposed method outperforms existing state-of-the-art methods across four datasets.
Graph neural networks (GNNs) and heterogeneous graph neural networks (HGNNs) are prominent techniques for homogeneous and heterogeneous graph representation learning, yet their performance in an end-to-end supervised framework greatly depends on the availability of task-specific supervision. To reduce the labeling cost, pre-training on self-supervised pretext tasks has become a popular paradigm, but there is often a gap between the pre-trained model and downstream tasks, stemming from the divergence in their objectives. To bridge the gap, prompt learning has risen as a promising direction especially in few-shot settings, without the need to fully fine-tune the pre-trained model. While there has been some early exploration of prompt-based learning on graphs, they primarily deal with homogeneous graphs, ignoring the heterogeneous graphs that are prevalent in downstream applications. In this paper, we propose HGPROMPT, a novel pre-training and prompting framework to unify not only pre-training and downstream tasks but also homogeneous and heterogeneous graphs via a dual-template design. Moreover, we propose dual-prompt in HGPROMPT to assist a downstream task in locating the most relevant prior to bridge the gaps caused by not only feature variations but also heterogeneity differences across tasks. Finally, we thoroughly evaluate and analyze HGPROMPT through extensive experiments on three public datasets.
In numerous real-world applications, it is quite common that sample information is partially available for some views due to machine breakdown or sensor failure, causing the problem of incomplete multi-view clustering (IMVC). While several IMVC approaches using view-shared anchors have successfully achieved pleasing performance improvement, (1) they generally construct anchors with only one dimension, which could deteriorate the multi-view diversity, bringing about serious information loss; (2) the constructed anchors are typically with a single size, which could not sufficiently characterize the distribution of the whole samples, leading to limited clustering performance. For generating view-shared anchors with multi-dimension and multi-size for IMVC, we design a novel framework called Diverse View-Shared Anchors based Incomplete multi-view clustering (DVSAI). Concretely, we associate each partial view with several potential spaces. In each space, we enable anchors to communicate among views and generate the view-shared anchors with space-specific dimension and size. Consequently, spaces with various scales make the generated view-shared anchors enjoy diverse dimensions and sizes. Subsequently, we devise an integration scheme with linear computational and memory expenditures to integrate the outputted multi-scale unified anchor graphs such that running spectral algorithm generates the spectral embedding. Afterwards, we theoretically demonstrate that DVSAI owns linear time and space costs, thus well-suited for tackling large-size datasets. Finally, comprehensive experiments confirm the effectiveness and advantages of DVSAI.
Multi-view graph clustering (MVGC) derives encouraging grouping results by seamlessly integrating abundant information inside heterogeneous data, and has captured surging focus recently. Nevertheless, the majority of current MVGC works involve at least one hyper-parameter, which not only requires additional efforts for tuning, but also leads to a complicated solving procedure, largely harming the flexibility and scalability of corresponding algorithms. To this end, in the article we are devoted to getting rid of hyper-parameters, and devise a non-parametric graph clustering (NpGC) framework to more practically partition multi-view data. To be specific, we hold that hyper-parameters play a role in balancing error item and regularization item so as to form high-quality clustering representations. Therefore, under without the assistance of hyper-parameters, how to acquire high-quality representations becomes the key. Inspired by this, we adopt two types of anchors, view-related and view-unrelated, to concurrently mine exclusive characteristics and common characteristics among views. Then, all anchors' information is gathered together via a consensus bipartite graph. By such ways, NpGC extracts both complementary and consistent multi-view features, thereby obtaining superior clustering results. Also, linear complexities enable it to handle datasets with over 120000 samples. Numerous experiments reveal NpGC's strong points compared to lots of classical approaches.
In recent years, graph-based fraud detection methods have garnered increasing attention for their superior ability to tackle the issue of camouflage in fraudulent scenarios. However, these methods often rely on a substantial proportion of samples as the training set, disregarding the reality of scarce annotated samples in real-life scenarios. As a theoretical framework within semi-supervised learning, the principle of consistency regularization posits that unlabeled samples should be classified into the same category as their own perturbations. Inspired by this principle, this study incorporates unlabeled samples as an auxiliary during model training, designing a novel barely supervised learning method to address the challenge of limited annotated samples in fraud detection. Specifically, to tackle the issue of camouflage in fraudulent scenarios, we employ disentangled representation learning based on edge information for a small subset of annotated nodes. This approach partitions node features into three distinct components representing different connected edges, providing a foundation for the subsequent augmentation of unlabeled samples. For the unlabeled nodes used in auxiliary training, we apply both strong and weak augmentation and design regularization losses to enhance the detection performance of the model in the context of extremely limited labeled samples. Across five publicly available datasets, the proposed model showcases its superior detection capability over baseline models.
Deep reinforcement learning (DRL) has gained immense success in many applications, including gaming AI, robotics, and system scheduling. Distributed algorithms and architectures have been vastly proposed (e.g., actor-learner architecture) to accelerate DRL training with large-scale server-based clusters. However, training on-policy algorithms with the actor-learner architecture unavoidably induces resource wasting due to synchronization between learners and actors, thus resulting in significantly extra billing. As a promising alternative, serverless computing naturally fits on-policy synchronization and alleviates resource wasting in distributed DRL training with pay-as-you-go pricing. Yet, none has leveraged serverless computing to facilitate DRL training. This paper proposes MinionsRL, the first serverless distributed DRL training framework that aims to accelerate DRL training- and cost-efficiency with dynamic actor scaling. We prototype MinionsRL on top of Microsoft Azure Container Instances and evaluate it with popular DRL tasks from OpenAI Gym. Extensive experiments show that MinionsRL reduces total training time by up to 52% and training cost by 86% compared to latest solutions.
Split learning is a computing resource-friendly distributed learning framework that protects client training data by splitting the model between the client and server. Previous work has proved that split learning faces a severe risk of privacy leakage, as a malicious server can recover the client's private data by hijacking the training process. In this paper, we first explore the vulnerability of split learning to server-side backdoor attacks, where our goal is to compromise the model's integrity. Since the server-side attacker cannot access the training data and client model in split learning, the traditional poisoning-based backdoor attack methods are no longer applicable. Therefore, constructing backdoor attacks in split learning poses significant challenges. Our strategy involves the attacker establishing a shadow model on the server side that can encode backdoor samples and guiding the client model to learn from this model during the training process, thereby enabling the client to acquire the same capability. Based on these insights, we propose a three-stage backdoor attack framework named SFI. Our attack framework minimizes assumptions about the attacker's background knowledge and ensures that the attack process remains imperceptible to the client. We implement SFI on various benchmark datasets, and extensive experimental results demonstrate its effectiveness and generality. For example, success rates of our attack on MNIST, Fashion, and CIFAR10 datasets all exceed 90%, with limited impact on the main task.
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.
Risk-Conditioned Reinforcement Learning: A Generalized Approach for Adapting to Varying Risk Measures
PDF ↗In application domains requiring mission-critical decision making, such as finance and robotics, the optimal policy derived by reinforcement learning (RL) often hinges on a preference for risk management. Yet, the dynamic nature of risk measures poses considerable challenges to achieving generalization and adaptation of risk-sensitive policies in the context of RL. In this paper, we propose a risk-conditioned RL model that enables rapid policy adaptation to varying risk measures via a unified risk representation, the Weighted Value-at-Risk (WV@R). To sample risk measures that avoid undue optimism, we construct a risk proposal network employing a conditional adversarial auto-encoder and a normalizing flow. This network establishes coherent representations for risk measures, preserving the continuity in terms of the Wasserstein distance on the risk measures. The normalizing flow is used to support non-crossing quantile regression that obtains valid samples for risk measures, and it is also applied to the agent’s critic to ascertain the preservation of monotonicity in quantile estimations. Through experiments with locomotion, finance, and self-driving scenarios, we show that our model is capable of adapting to a range of risk measures, achieving comparable performance to the baseline models individually trained for each measure. Our model often outperforms the baselines, especially in the cases when exploration is required during training but risk-aversion is favored during evaluation.
In recent years, many methods have been proposed to address the zero-shot sketch-based image retrieval (ZS-SBIR) task, which is a practical problem in many applications. However, in real-world scenarios, on the one hand, we can not obtain training data with the same distribution as the test data, and on the other hand, the labels of training data are not available as usual. To tackle this issue, we focus on a new problem, namely unsupervised zero-shot sketch-based image retrieval (UZS-SBIR), where the available training data does not have labels while the training and testing categories are not overlapping. In this paper, we introduce a new asymmetric mutual alignment method (AMA) including a self-distillation module and a cross-modality mutual alignment module. First, we conduct self-distillation to extract the feature embeddings from unlabeled data. Due to the lack of available information in an unsupervised manner, we employ the cross-modality mutual alignment module to further excavate underlying intra-modality and inter-modality relationships from unlabeled data, and take full advantage of these correlations to align the feature embeddings in image and sketch domains. Meanwhile, the feature representations are enhanced by the intra-modality clustering relations, leading to better generalization ability to unseen classes. Moreover, we conduct an asymmetric strategy to update the teacher and student networks, respectively. Extensive experimental results on several benchmark datasets demonstrate the superiority of our method.
The integration of Spiking Neural Networks (SNNs) and Graph Neural Networks (GNNs) is gradually attracting attention due to the low power consumption and high efficiency in processing the non-Euclidean data represented by graphs. However, as a common problem, dynamic graph representation learning faces challenges such as high complexity and large memory overheads. Current work often uses SNNs instead of Recurrent Neural Networks (RNNs) by using binary features instead of continuous ones for efficient training, which overlooks graph structure information and leads to the loss of details during propagation. Additionally, optimizing dynamic spiking models typically requires the propagation of information across time steps, which increases memory requirements. To address these challenges, we present a framework named Dynamic Spiking Graph Neural Networks (Dy-SIGN). To mitigate the information loss problem, Dy-SIGN propagates early-layer information directly to the last layer for information compensation. To accommodate the memory requirements, we apply the implicit differentiation on the equilibrium state, which does not rely on the exact reverse of the forward computation. While traditional implicit differentiation methods are usually used for static situations, Dy-SIGN extends it to the dynamic graph setting. Extensive experiments on three large-scale real-world dynamic graph datasets validate the effectiveness of Dy-SIGN on dynamic node classification tasks with lower computational costs.
Capturing the underlying structural causal relations represented by Directed Acyclic Graphs (DAGs) has been a fundamental task in various AI disciplines. Causal DAG learning via the continuous optimization framework has recently achieved promising performance in terms of accuracy and efficiency. However, most methods make strong assumptions of homoscedastic noise, i.e., exogenous noises have equal variances across variables, observations, or even both. The noises in real data usually violate both assumptions due to the biases introduced by different data collection processes. To address the heteroscedastic noise issue, we introduce relaxed implementable sufficient conditions and prove the identifiability of a general class of SEM subject to those conditions. Based on the identifiable general SEM, we propose a novel formulation for DAG learning which accounts for the noise variance variation across variables and observations. We then propose an effective two-phase iterative DAG learning algorithm to address the increasing optimization difficulties and learn a causal DAG from data with heteroscedastic variables noise under varying variance. We show significant empirical gains of the proposed approaches over state-of-the-art methods on both synthetic data and real data.
Multi-view K-means clustering successfully generalizes K-means from single-view to multi-view, and obtains excellent clustering performance. In every view, it makes each data point close to the center of the corresponding cluster. However, multi-view K-means only considers the compactness of each cluster, but ignores the separability of different clusters, which is of great importance to producing a good clustering result. In this paper, we propose Discriminatively Fuzzy Multi-view K-means clustering with Local Structure Preserving (DFMKLS). On the basis of minimizing the distance between each data point and the center of the corresponding cluster, DFMKLS separates clusters by maximizing the distance between the centers of pairwise clusters. DFMKLS also relaxes its objective by introducing the idea of fuzzy clustering, which calculates the probability that a data point belongs to each cluster. Considering multi-view K-means mainly focuses on the global information of the data, to efficiently use the local information, we integrate the local structure preserving into the framework of DFMKLS. The effectiveness of DFMKLS is evaluated on benchmark multi-view datasets. It obtains superior performances than state-of-the-art multi-view clustering methods, including multi-view K-means.
Near-Optimal Resilient Aggregation Rules for Distributed Learning Using 1-Center and 1-Mean Clustering with Outliers
PDF ↗Byzantine machine learning has garnered considerable attention in light of the unpredictable faults that can occur in large-scale distributed learning systems. The key to secure resilience against Byzantine machines in distributed learning is resilient aggregation mechanisms. Although abundant resilient aggregation rules have been proposed, they are designed in ad-hoc manners, imposing extra barriers on comparing, analyzing, and improving the rules across performance criteria. This paper studies near-optimal aggregation rules using clustering in the presence of outliers. Our outlier-robust clustering approach utilizes geometric properties of the update vectors provided by workers. Our analysis show that constant approximations to the 1-center and 1-mean clustering problems with outliers provide near-optimal resilient aggregators for metric-based criteria, which have been proven to be crucial in the homogeneous and heterogeneous cases respectively. In addition, we discuss two contradicting types of attacks under which no single aggregation rule is guaranteed to improve upon the naive average. Based on the discussion, we propose a two-phase resilient aggregation framework. We run experiments for image classification using a non-convex loss function. The proposed algorithms outperform previously known aggregation rules by a large margin with both homogeneous and heterogeneous data distributions among non-faulty workers. Code and appendix are available at https://github.com/jerry907/AAAI24-RASHB.
The Evidential Regression Network (ERN) represents a novel approach that integrates deep learning with Dempster-Shafer's theory to predict a target and quantify the associated uncertainty. Guided by the underlying theory, specific activation functions must be employed to enforce non-negative values, which is a constraint that compromises model performance by limiting its ability to learn from all samples. This paper provides a theoretical analysis of this limitation and introduces an improvement to overcome it. Initially, we define the region where the models can't effectively learn from the samples. Following this, we thoroughly analyze the ERN and investigate this constraint. Leveraging the insights from our analysis, we address the limitation by introducing a novel regularization term that empowers the ERN to learn from the whole training set. Our extensive experiments substantiate our theoretical findings and demonstrate the effectiveness of the proposed solution.
Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster
PDF ↗Human brains can continually acquire and learn new skills and knowledge over time from a dynamically changing environment without forgetting previously learnt information. Such a capacity can selectively transfer some important and recently seen information to the persistent knowledge regions of the brain. Inspired by this intuition, we propose a new memory-based approach for image reconstruction and generation in continual learning, consisting of a temporary and evolving memory, with two different storage strategies, corresponding to the temporary and permanent memorisation. The temporary memory aims to preserve up-to-date information while the evolving memory can dynamically increase its capacity in order to preserve permanent knowledge information. This is achieved by the proposed memory expansion mechanism that selectively transfers those data samples deemed as important from the temporary memory to new clusters defined within the evolved memory according to an information novelty criterion. Such a mechanism promotes the knowledge diversity among clusters in the evolved memory, resulting in capturing more diverse information by using a compact memory capacity. Furthermore, we propose a two-step optimization strategy for training a Variational Autoencoder (VAE) to implement generation and representation learning tasks, which updates the generator and inference models separately using two optimisation paths. This approach leads to a better trade-off between generation and reconstruction performance. We show empirically and theoretically that the proposed approach can learn meaningful latent representations while generating diverse images from different domains. The source code and supplementary material (SM) are available at https://github.com/dtuzi123/DEMC.
Vision Transformers (ViTs) represent self-attention-based network backbones shown to be efficient in many individual tasks, but which have not been explored in Task-Free Continual Learning (TFCL) so far. Most existing ViT-based approaches for Continual Learning (CL) are relying on task information. In this study, we explore the advantages of the ViT in a more challenging CL scenario where the task boundaries are unavailable during training. To address this learning paradigm, we propose the Task-Free Dynamic Sparse Vision Transformer (TFDSViT), which can dynamically build new sparse experts, where each expert leverages sparsity to allocate the model's capacity for capturing different information categories over time. To avoid forgetting and ensure efficiency in reusing the previously learned knowledge in subsequent learning, we propose a new dynamic dual attention mechanism consisting of the Sparse Attention (SA') and Knowledge Transfer Attention (KTA) modules. The SA' refrains from updating some previously learned attention blocks for preserving prior knowledge. The KTA uses and regulates the information flow of all previously learned experts for learning new patterns. The proposed dual attention mechanism can simultaneously relieve forgetting and promote knowledge transfer for a dynamic expansion model in a task-free manner. We also propose an energy-based dynamic expansion mechanism using the energy as a measure of novelty for the incoming samples which provides appropriate expansion signals leading to a compact network architecture for TFDSViT. Extensive empirical studies demonstrate the effectiveness of TFDSViT. The code and supplementary material (SM) are available at https://github.com/dtuzi123/TFDSViT.