Diffusion-based models have been recently shown to be high-quality data generators. However, their performance severely degrades when training on non-stationary changing data distributions in an online manner, due to the catastrophic forgetting. In this paper, we propose enabling the diffusion model with a novel Dynamic Expansion Memory Unit (DEMU) methodology that adaptively creates new memory buffers, to be added to a memory system, in order to preserve information deemed critical for training the model. Having a selective memory unit is essential for training diffusion networks, which are expensive to train, especially when deployed in resource-constrained environments. A Maximum Mean Discrepancy (MMD) based expansion mechanism, that evaluates probabilistic distances between each of the previously defined memory buffers and the newly given data, and uses them as expansion signals, is employed for ensuring the diversity of information learning. We propose a new model expansion mechanism to automatically add new diffusion models as experts in a mixture system, which enhances the multi-domain image generation performance. Also a novel memory compaction approach is proposed to automatically remove statistically overlapping memory units, through a graph relationship evaluation, preventing the limitless expansion of DEMU. Comprehensive results show that the proposed approach performs better than the state-of-the-art.
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
The diffusion model has lately been shown to achieve remarkable performances through its ability of generating high quality images. However, current diffusion model studies consider only learning from a single data distribution, resulting in catastrophic forgetting when attempting to learn new data. In this paper, we explore a more realistic learning scenario where training data is continuously acquired. We propose the Dynamic Expansion Diffusion Model (DEDM) for addressing catastrophic forgetting and data distribution shifts under Online Task-Free Continual Learning (OTFCL) paradigm. New diffusion components are added to a mixture model following the evaluation of a criterion which compares the probabilistic representation of the new data with the existing knowledge of the DEDM model. In addition, to maintain an optimal architecture, we propose a component discovery approach that ensures the diversity of knowledge while minimizing the total number of parameters in the DEDM. Furthermore, we show how the proposed DEDM can be implemented as a teacher module in a unified framework for representation learning. In this approach, knowledge distillation is proposed for training a student module aiming to compress the teacher's knowledge into the latent space of the student.
Lately, deep generative models have achieved excellent results after learning pre-defined and static data distribution. Meanwhile, their performance on continual learning suffers from degeneration, caused by catastrophic forgetting. In this paper, we study the unsupervised generative modelling in a more realistic continual learning scenario, where class and task information are absent during both training and inference learning phases. To implement this goal, the proposed memory approach consists of a temporary memory system, which stores data examples while a dynamic expansion memory system would gradually preserve those samples that are crucial for long-term memorization. A novel memory expansion mechanism is then proposed, by employing optimal transport distances between the statistics of memorized samples and each newly seen datum. This paper proposes the Sinkhorn-based Dual Dynamic Memory (SDDM) method, by considering Sinkhorn distance as an optimal transport measure, for evaluating the significance of the data to be stored in the memory buffer. The Sinkhorn transport algorithm leads to preserving a diversity of samples within a compact memory capacity. The memory buffering approach does not interact with the model's training process and can be optimized independently in both supervised and unsupervised learning without any modifications. Moreover, we also propose a novel dynamic model expansion mechanism to automatically increase the model's capacity whenever necessary, which can deal with infinite data streams and further improve the model's performance. Experimental results show that the proposed approach achieves state-of-the-art performance in both supervised and unsupervised learning.
Pre-trained Vision Mamba~(Vim) models have demonstrated exceptional performance across various computer vision tasks in a computationally efficient manner, attributed to their unique design of selective state space models. To further extend their applicability to diverse downstream vision tasks, Vim models can be adapted using the efficient fine-tuning technique known as visual prompting. However, existing visual prompting methods are predominantly tailored for Vision Transformer (ViT)-based models that leverage global attention, neglecting the distinctive sequential token-wise compression and propagation characteristics of Vim. Specifically, existing prompt tokens prefixed to the sequence are insufficient to effectively activate the input and forget gates across the entire sequence, hindering the extraction and propagation of discriminative information. To address this limitation, we introduce a novel Selective Visual Prompting (SVP) method specifically for the efficient fine-tuning of Vim. To prevent the loss of discriminative information during state space propagation, SVP employs lightweight selective prompters for token-wise prompt generation, ensuring adaptive activation of the update and forget gates within Mamba blocks to promote discriminative information propagation. Moreover, considering that Vim propagates both shared cross-layer information and specific inner-layer information, we further refine SVP with a dual-path structure: Cross-Prompting and Inner-Prompting. Cross-Prompting utilizes shared parameters across layers, while Inner-Prompting employs distinct parameters, promoting the propagation of both shared and specific information, respectively. Extensive experimental results on various large-scale benchmarks demonstrate that our proposed SVP significantly outperforms state-of-the-art methods.
ScaleOT: Privacy-utility-scalable Offsite-tuning with Dynamic LayerReplace and Selective Rank Compression
PDF ↗Offsite-tuning is a privacy-preserving method for tuning large language models (LLMs) by sharing a lossy compressed emulator from the LLM owners with data owners for downstream task tuning. This approach protects the privacy of both the model and data owners. However, current offsite tuning methods often suffer from adaptation degradation, high computational costs, and limited protection strength due to uniformly dropping LLM layers or relying on expensive knowledge distillation. To address these issues, we propose ScaleOT, a novel privacy-utility-scalable offsite-tuning framework that effectively balances privacy and utility. ScaleOT introduces a novel layerwise lossy compression algorithm that uses reinforcement learning to obtain the importance of each layer. It employs lightweight networks, termed harmonizers, to replace the raw LLM layers. By combining important original LLM layers and harmonizers in different ratios, ScaleOT generates emulators tailored for optimal performance with various model scales for enhanced privacy protection. Additionally, we present a rank reduction method to further compress the original LLM layers, significantly enhancing privacy with negligible impact on utility. Comprehensive experiments show that ScaleOT can achieve nearly lossless offsite tuning performance compared with full fine-tuning while obtaining better model privacy.
In this paper, we address a practical distributed Bayesian learning problem with asynchronous measurements and predictions due to diverse computational conditions. To this end, asynchronous distributed Gaussian process (AsyncDGP) regression is proposed, which is the first effective online distributed Gaussian processes (GPs) approach to improve the prediction accuracy in real-time learning tasks. By leveraging the devised evaluation criterion and established prediction error bounds, AsyncDGP enables the distinction of contributions of each model for prediction ensembling using aggregation strategy. Furthermore, we extend its utility to dynamic systems by introducing a learning-based control law, ensuring guaranteed control performance in safety-critical applications. Additionally, a networked online learning simulation platform for distributed GPs, namely online GP gym (GPgym), is introduced for testing the performance of learning and control of dynamical systems. Numerical simulations within GPgym across regression tasks with real-world data sets and dynamical control scenarios demonstrate the effectiveness and applicability of AsyncDGP.
AutoMMLab: Automatically Generating Deployable Models from Language Instructions for Computer Vision Tasks
PDF ↗Automated machine learning (AutoML) is a collection of techniques designed to automate the machine learning development process. While traditional AutoML approaches have been successfully applied in several critical steps of model development (e.g. hyperparameter optimization), there lacks a AutoML system that automates the entire end-to-end model production workflow for computer vision. To fill this blank, we propose a novel request-to-model task, which involves understanding the user's natural language request and execute the entire workflow to output production-ready models. This empowers non-expert individuals to easily build task-specific models via a user-friendly language interface. To facilitate development and evaluation, we develop a new experimental platform called AutoMMLab and a new benchmark called LAMP for studying key components in the end-to-end request-to-model pipeline. Hyperparameter optimization (HPO) is one of the most important components for AutoML. Traditional approaches mostly rely on trial-and-error, leading to inefficient parameter search. To solve this problem, we propose a novel LLM-based HPO algorithm, called HPO-LLaMA. Equipped with extensive knowledge and experience in model hyperparameter tuning, HPO-LLaMA achieves significant improvement of HPO efficiency.
We investigate safe online convex optimization (SOCO), where each decision must satisfy a set of unknown linear constraints. Assuming that the unknown constraints can be observed with a sub-Gaussian noise for each chosen decision, previous studies have established a high-probability regret bound of O(T^{2/3}). However, this assumption may not hold in many practical scenarios. To address this limitation, in this paper, we relax the assumption to allow any noise that admits finite (1+ε)-th moments for some ε∈(0,1], and propose two algorithms that enjoy an O(T^{c_ε}) regret bound with high probability, where T is the time horizon and c_ε=(1+ε)/(1+2ε). The key idea of our two algorithms is to respectively utilize the median-of-means and truncation techniques to achieve accurate estimation under heavy-tailed noises. To the best of our knowledge, these are the first algorithms designed to handle SOCO with heavy-tailed observation noises.
Multi-instance partial-label learning (MIPL) is a paradigm where each training example is encapsulated as a multi-instance bag associated with the candidate label set, which includes one true label and several false positives. Current MIPL algorithms typically assume that all instances are independent, thereby neglecting the dependencies and heterogeneity inherent in MIPL data. Moreover, these algorithms often prove to be excessively time-consuming when dealing with complex datasets, significantly limiting the practical application of MIPL. In this paper, we propose FastMIPL, a framework that employs mixed-effects model to explicitly capture the dependencies and heterogeneity among instances and bags. FastMIPL is able to learn from MIPL data both effectively and efficiently by utilizing the predefined dependencies modeling module and leveraging the posterior predictive probability disambiguation strategy. Experiments show that the performance of FastMIPL is highly competitive to state-of-the-art methods, while significantly reducing computational time in benchmark and the real-world datasets.
Federated bilevel optimization (FBO) has garnered significant attention lately, driven by its promising applications in meta-learning and hyperparameter optimization. Existing algorithms generally aim to approximate the gradient of the upper-level objective function (hypergradient) in the federated setting. However, because of the nonlinearity of the hypergradient and client drift, they often involve complicated computations. These computations, like multiple optimization sub-loops and second-order derivative evaluations, end up with significant memory consumption and high computational costs. In this paper, we propose a computationally and memory-efficient FBO algorithm named MemFBO. MemFBO features a fully single-loop structure with all involved variables updated simultaneously, and uses only first-order gradient information for all local updates. We show that MemFBO exhibits a linear convergence speedup with milder assumptions in both partial and full client participation scenarios. We further implement MemFBO in a novel FBO application for federated data cleaning. Our experiments, conducted on this application and federated hyper-representation, demonstrate the effectiveness of the proposed algorithm.
Conditional independence (CI) testing is a fundamental task in modern statistics and machine learning. The conditional randomization test (CRT) was recently introduced to test whether two random variables, X and Y, are conditionally independent given a potentially high-dimensional set of random variables, Z. The CRT operates exceptionally well under the assumption that the conditional distribution X|Z is known. However, since this distribution is typically unknown in practice, accurately approximating it becomes crucial. In this paper, we propose using conditional diffusion models (CDMs) to learn the distribution of X|Z. Theoretically and empirically, it is shown that CDMs closely approximate the true conditional distribution. Furthermore, CDMs offer a more accurate approximation of X|Z compared to GANs, potentially leading to a CRT that performs better than those based on GANs. To accommodate complex dependency structures, we utilize a computationally efficient classifier-based conditional mutual information (CMI) estimator as our test statistic. The proposed testing procedure performs effectively without requiring assumptions about specific distribution forms or feature dependencies, and is capable of handling mixed-type conditioning sets that include both continuous and discrete variables. Theoretical analysis shows that our proposed test achieves a valid control of the type I error. A series of experiments on synthetic data demonstrates that our new test effectively controls both type-I and type-II errors, even in high dimensional scenarios.
Parameter-efficient fine-tuning (PEFT) has been widely employed for domain adaptation, with LoRA being one of the most prominent methods due to its simplicity and effectiveness. However, in multi-task learning (MTL) scenarios, LoRA tends to obscure the distinction between tasks by projecting sparse high-dimensional features from different tasks into the same dense low-dimensional intrinsic space. This leads to task interference and suboptimal performance for LoRA and its variants. To tackle this challenge, we propose MTL-LoRA, which retains the advantages of low-rank adaptation while significantly enhancing MTL capabilities. MTL-LoRA augments LoRA by incorporating additional task-adaptive parameters that differentiate task-specific information and capture shared knowledge across various tasks within low-dimensional spaces. This approach enables pretrained models to jointly adapt to different target domains with a limited number of trainable parameters. Comprehensive experimental results, including evaluations on public academic benchmarks for natural language understanding, commonsense reasoning, and image-text understanding, as well as real-world industrial text Ads relevance datasets, demonstrate that MTL-LoRA outperforms LoRA and its various variants with comparable or even fewer learnable parameters in MTL setting.
Anomaly detection aims to identify deviations from normal patterns within data. This task is particularly crucial in dynamic graphs, which are common in applications like social networks and cybersecurity, due to their evolving structures and complex relationships. Although recent deep learning-based methods have shown promising results in anomaly detection on dynamic graphs, they often lack of generalizability. In this study, we propose GeneralDyG, a method that samples temporal ego-graphs and sequentially extracts structural and temporal features to address the three key challenges in achieving generalizability: Data Diversity, Dynamic Feature Capture, and Computational Cost. Extensive experimental results demonstrate that our proposed GeneralDyG significantly outperforms state-of-the-art methods on four real-world datasets.
We investigate the online nonsubmodular optimization with delayed feedback in the bandit setting, where the loss function is α-weakly DR-submodular and β-weakly DR-supermodular. Previous work has established an (α,β)-regret bound of O(nd^⅓T^⅔), where n is the dimensionality and d is the maximum delay. However, its regret bound relies on the maximum delay and is thus sensitive to irregular delays. Additionally, it couples the effects of delays and bandit feedback as its bound is the product of the delay term and the O(nT^⅔) regret bound in the bandit setting without delayed feedback. In this paper, we develop two algorithms to address these limitations, respectively. Firstly, we propose a novel method, namely DBGD-NF, which employs the one-point gradient estimator and utilizes all the available estimated gradients in each round to update the decision. It achieves a better O(nd̅^⅓T^⅔) regret bound, which is relevant to the average delay d̅ = 1/T ∑ₜ₌₁ᵀ dₜ
As an essential technique for Graph Contrastive Learning (GCL), Graph Augmentation (GA) improves the generalization capability of the GCLs by introducing different forms of the same graph. To ensure information integrity, existing GA strategies have been designed to simultaneously process the two types of information available in graphs: node attributes and graph topology. Nonetheless, these strategies tend to augment the two types of graph information separately, ignoring their correlation, resulting in limited representation ability. To overcome this drawback, this paper proposes a novel GCL framework with a Joint spectrAl augMentation, named GCL-JAM. Motivated the equivalence between the graph learning objective on an attribute graph and the spectral clustering objective on the attribute-interpolated graph, the node attributes are first abstracted as another type of node to harmonize the node attributes and graph topology. The newly constructed graph is then utilized to perform spectral augmentation to capture the correlation during augmentation. Theoretically, the proposed joint spectral augmentation is proved to perturb more inter-class edges and noise attributes compared to separate augmentation methods. Extensive experiments on homophily and heterophily graphs validate the effectiveness and universality of GCL-JAM.
Contrastive Language-Image Pre-training (CLIP) has achieved excellent performance over a wide range of tasks. However, the effectiveness of CLIP heavily relies on a substantial corpus of pre-training data, resulting in notable consumption of computational resources. Although knowledge distillation has been widely applied in single modality models, how to efficiently expand knowledge distillation to vision-language foundation models with extensive data remains relatively unexplored. In this paper, we introduce CLIP-CID, a novel distillation mechanism that effectively transfers knowledge from a large vision-language foundation model to a smaller model. We initially propose a simple but efficient image semantic balance method to reduce transfer learning bias and improve distillation efficiency. This method filters out 43.7% of image-text pairs from the LAION400M while maintaining superior performance. After that, we leverage cluster-instance discrimination to facilitate knowledge transfer from the teacher model to the student model, thereby empowering the student model to acquire a holistic semantic comprehension of the pre-training data. Experimental results demonstrate that CLIP-CID achieves state-of-the-art performance on various downstream tasks including linear probe and zero-shot classification.
A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset
PDF ↗Recently, the advent of generative AI technologies has made transformational impacts on our daily lives, yet its application in scientific applications remains in its early stages. Data scarcity is a major, well-known barrier in data-driven scientific computing, so physics-guided generative AI holds significant promise. In scientific computing, most tasks study the conversion of multiple data modalities to describe physical phenomena, for example, spatial and waveform in seismic imaging, time and frequency in signal processing, and temporal and spectral in climate modeling; as such, multi-modal pairwise data generation is highly required instead of single-modal data generation, which is usually used in natural images (e.g., faces, scenery). Moreover, in real-world applications, the unbalance of available data in terms of modalities commonly exists; for example, the spatial data (i.e., velocity maps) in seismic imaging can be easily simulated, but real-world seismic waveform is largely lacking. While the most recent efforts enable the powerful diffusion model to generate multi-modal data, how to leverage the unbalanced available data is still unclear. In this work, we use seismic imaging in subsurface geophysics as a vehicle to present "UB-Diff", a novel diffusion model for multi-modal paired scientific data generation. One major innovation is a one-in-two-out encoder-decoder network structure, which can ensure pairwise data is obtained from a co-latent representation. Then, the co-latent representation will be used by the diffusion process for pairwise data generation. Experimental results on the OpenFWI dataset show that UB-Diff significantly outperforms existing techniques in terms of Fréchet Inception Distance (FID) score and pairwise evaluation, indicating the generation of reliable and useful multi-modal pairwise data.
Split Federated Learning (SFL) splits and collaboratively trains a shared model between clients and server, where clients transmit activations and client-side models to server for updates. Recent SFL studies assume synchronous transmission of activations and client-side models from clients to server. However, due to significant variations in computational and communication capabilities among clients, activations and client-side models arrive at server asynchronously. The delay caused by asynchrony significantly degrades the performance of SFL. To address this issue, we consider an asynchronous SFL framework, where an activation buffer and a model buffer are embedded on the server to manage the asynchronously transmitted activations and client-side models, respectively. Furthermore, as asynchronous activation transmissions cause the buffer to frequently receive activations from resource-rich clients, leading to biased updates of the server-side model, we propose Generative activations-aided Asynchronous SFL (GAS). In GAS, the server maintains an activation distribution for each label based on received activations and generates activations from these distributions according to the degree of bias. These generative activations are then used to assist in updating the server-side model, ensuring more accurate updates. We derive a tighter convergence bound, and our experiments demonstrate the effectiveness of the proposed method.
Acting Beyond Learning: Imagination-Assisted Decision-Making in the Visual-based Multi-Agent Cooperative Scenarios
PDF ↗Learning optimal policies in multi-agent cooperative settings with visual observations is significant and challenging. Agents must first perform state representation learning for their image observations and then learn policies in the abstracted state space. Aiming at this problem, we propose a novel model-based MARL method named Contrastive Latent World for Policy Optimization (CLWPO). In CLWPO, we first design a state representation model to facilitate learning in the latent state space. With the support of this model, we construct the latent world and introduce a contrastive variational bound (CVB) to optimize it. Subsequently, we develop a heuristic policy optimization (HPO) scheme, incorporating model-free learning with model-based planning to obtain robust policies that predict future behaviors. In particular, in the planning, we maintain a queue of teammate models and calculate an adaptive rollout length for each agent to support their self-imagination and reduce the model-based return discrepancy. Finally, we conducted extensive experiments in the PettingZoo benchmark, and results show that CLWPO significantly enhances learning efficiency and improves agent performance compared to state-of-the-art MARL methods.
The optimization-based meta-learning approach is gaining increased traction because of its unique ability to quickly adapt to a new task using only small amounts of data. However, existing optimization-based meta-learning approaches, such as MAML, ANIL and their variants, generally employ backpropagation for upper-level gradient estimation, which requires using historical lower-level parameters/gradients and thus increases computational and memory overhead in each iteration. In this paper, we propose a meta-learning algorithm that can avoid using historical parameters/gradients and significantly reduce memory costs in each iteration compared to existing optimization-based meta-learning approaches. In addition to memory reduction, we prove that our proposed algorithm converges sublinearly with the iteration number of upper-level optimization, and the convergence error decays sublinearly with the batch size of sampled tasks. In the specific case in terms of deterministic meta-learning, we also prove that our proposed algorithm converges to an exact solution. Moreover, we quantify the computational complexity of the algorithm, which matches existing convergence results on meta-learning even without using any historical parameters/gradients. Experimental results on meta-learning benchmarks confirm the efficacy of our proposed algorithm.