ICLR 2025 Papers

High-quality open-source datasets, which necessitate substantial efforts for curation, has become the primary catalyst for the swift progress of deep learning. Concurrently, protecting these datasets is paramount for the well-being of the data owner. Dataset ownership verification emerges as a crucial method in this domain, but existing approaches are often limited to supervised models and cannot be directly extended to increasingly popular unsupervised pre-trained models. In this work, we propose the first dataset ownership verification method tailored specifically for self-supervised pre-trained models by contrastive learning. Its primary objective is to ascertain whether a suspicious black-box backbone has been pre-trained on a specific unlabeled dataset, aiding dataset owners in upholding their rights. The proposed approach is motivated by our empirical insights that when models are trained with the target dataset, the unary and binary instance relationships within the embedding space exhibit significant variations compared to models trained without the target dataset. We validate the efficacy of this approach across multiple contrastive pre-trained models including SimCLR, BYOL, SimSiam, MOCO v3, and DINO. The results demonstrate that our method rejects the null hypothesis with a $p$-value markedly below $0.05$, surpassing all previous methodologies. Our code is available at https://github.com/xieyc99/DOV4CL.

Self-supervised learning (SSL) aims to learn meaningful representations from unlabeled data. Orthogonal Low-rank Embedding (OLE) shows promise for SSL by enhancing intra-class similarity in a low-rank subspace and promoting inter-class dissimilarity in a high-rank subspace, making it particularly suitable for multi-view learning tasks. However, directly applying OLE to SSL poses significant challenges: (1) the virtually infinite number of "classes" in SSL makes achieving the OLE objective impractical, leading to representational collapse; and (2) low-rank constraints may fail to distinguish between positively and negatively correlated features, further undermining learning. To address these issues, we propose SSOLE (Self-Supervised Orthogonal Low-rank Embedding), a novel framework that integrates OLE principles into SSL by (1) decoupling the low-rank and high-rank enforcement to align with SSL objectives; and (2) applying low-rank constraints to feature deviations from their mean, ensuring better alignment of positive pairs by accounting for the signs of cosine similarities. Our theoretical analysis and empirical results demonstrate that these adaptations are crucial to SSOLE’s effectiveness. Moreover, SSOLE achieves competitive performance across SSL benchmarks without relying on large batch sizes, memory banks, or dual-encoder architectures, making it an efficient and scalable solution for self-supervised tasks. Code is available at https://github.com/husthuaan/ssole.

In this work, we focus on Weakly Supervised Spatio-Temporal Video Grounding (WSTVG). It is a multimodal task aimed at localizing specific subjects spatio-temporally based on textual queries without bounding box supervision. Motivated by recent advancements in multi-modal foundation models for grounding tasks, we first explore the potential of state-of-the-art object detection models for WSTVG. Despite their robust zero-shot capabilities, our adaptation reveals significant limitations, including inconsistent temporal predictions, inadequate understanding of complex queries, and challenges in adapting to difficult scenarios. We propose CoSPaL (Contextual Self-Paced Learning), a novel approach which is designed to overcome these limitations. CoSPaL integrates three core components: (1) Tubelet Phrase Grounding (TPG), which introduces spatio-temporal prediction by linking textual queries to tubelets; (2) Contextual Referral Grounding (CRG), which improves comprehension of complex queries by extracting contextual information to refine object identification over time; and (3) Self-Paced Scene Understanding (SPS), a training paradigm that progressively increases task difficulty, enabling the model to adapt to complex scenarios by transitioning from coarse to fine-grained understanding.

Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms.

While conventional data are represented as discrete vectors, Implicit Neural Representations (INRs) utilize neural networks to represent data points as continuous functions. By incorporating a shared network that maps latent vectors to individual functions, one can model the distribution of functional data, which has proven effective in many applications, such as learning 3D shapes, surface reflectance, and operators. However, the infinite-dimensional nature of these representations makes them prone to overfitting, necessitating sufficient regularization. Naïve regularization methods -- those commonly used with discrete vector representations -- may enforce smoothness to increase robustness but result in a loss of data fidelity due to improper handling of function coordinates. To overcome these challenges, we start by interpreting the mapping from latent variables to INRs as a parametrization of a Riemannian manifold. We then recognize that preserving geometric quantities -- such as distances and angles -- between the latent space and the data manifold is crucial. As a result, we obtain a manifold with minimal intrinsic curvature, leading to robust representations while maintaining high-quality data fitting. Our experiments on various data modalities demonstrate that our method effectively discovers a well-structured latent space, leading to robust data representations even for challenging datasets, such as those that are small or noisy.

T cells are a key component of the adaptive immune system, targeting infections, cancers, and allergens with specificity encoded by their T cell receptors (TCRs), and retaining a memory of their targets. High-throughput TCR repertoire sequencing captures a cross-section of TCRs that encode the immune history of any subject, though the data are heterogeneous, high dimensional, sparse, and mostly unlabeled. Sets of TCRs responding to the same antigen, *i.e.*, a protein fragment, co-occur in subjects sharing immune genetics and exposure history. Here, we leverage TCR co-occurrence across a large set of TCR repertoires and employ the GloVe (Pennington et al., 2014) algorithm to derive low-dimensional, dense vector representations (embeddings) of TCRs. We then aggregate these TCR embeddings to generate subject-level embeddings based on observed *subject-specific* TCR subsets. Further, we leverage random projection theory to improve GloVe's computational efficiency in terms of memory usage and training time. Extensive experimental results show that TCR embeddings targeting the same pathogen have high cosine similarity, and subject-level embeddings encode both immune genetics and pathogenic exposure history.

Magnetic Resonance Imaging (MRI), including diffusion MRI (dMRI), serves as a ``microscope'' for anatomical structures and routinely mitigates the influence of low signal-to-noise ratio scans by compromising temporal or spatial resolution. However, these compromises fail to meet clinical demands for both efficiency and precision. Consequently, denoising is a vital preprocessing step, particularly for dMRI, where clean data is unavailable. In this paper, we introduce Di-Fusion, a fully self-supervised denoising method that leverages the latter diffusion steps and an adaptive sampling process. Unlike previous approaches, our single-stage framework achieves efficient and stable training without extra noise model training and offers adaptive and controllable results in the sampling process. Our thorough experiments on real and simulated data demonstrate that Di-Fusion achieves state-of-the-art performance in microstructure modeling, tractography tracking, and other downstream tasks. Code is available at https://github.com/FouierL/Di-Fusion.

The increasing scale of graph datasets significantly enhances deep learning models but also presents substantial training challenges. Graph dataset condensation has emerged to condense large datasets into smaller yet informative ones that maintain similar test performance. However, these methods require downstream usage to match the original dataset and task, which is impractical in real-world scenarios. Our empirical studies show that existing methods fail in "cross-task" and "cross-dataset" scenarios, often performing worse than training from scratch. To address these challenges, we propose a novel method: Self-supervised and Transferable Graph dataset Condensation (ST-GCond). For cross-task transferability, we propose a task-disentangled meta optimization strategy to adaptively update the condensed graph according to the task relevance, encouraging information preservation for various tasks. For cross-dataset transferability, we propose a multi-teacher self-supervised optimization strategy to incorporate auxiliary self-supervised tasks to inject universal knowledge into the condensed graph. Additionally, we incorporate mutual information guided joint condensation mitigating the potential conflicts and ensure the condensing stability. Experiments on both node-level and graph-level datasets show that ST-GCond outperforms existing methods by 2.5% to 18.7% in all cross-task and cross-dataset scenarios, and also achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.

Knowledge distillation leverages a teacher model to improve the training of a student model. A persistent challenge is that a better teacher does not always yield a better student, to which a common mitigation is to use additional supervision from several “intermediate” teachers. One empirically validated variant of this principle is progressive distillation, where the student learns from successive intermediate checkpoints of the teacher. Using sparse parity as a sandbox, we identify an implicit curriculum as one mechanism through which progressive distillation accelerates the student’s learning. This curriculum is available only through the intermediate checkpoints but not the final converged one, and imparts both empirical acceleration and a provable sample complexity benefit to the student. We then extend our investigation to Transformers trained on probabilistic context-free grammars (PCFGs) and real-world pre-training datasets (Wikipedia and Books). Through probing the teacher model, we identify an analogous implicit curriculum where the model progressively learns features that capture longer context. Our theoretical and empirical findings on sparse parity, complemented by empirical observations on more complex tasks, highlight the benefit of progressive distillation via implicit curriculum across setups.

In scenarios where language models must incorporate new information efficiently without extensive retraining, traditional fine-tuning methods are prone to overfitting, degraded generalization, and unnatural language generation. To address these limitations, we introduce Consistent In-Context Editing (ICE), a novel approach leveraging the model's in-context learning capability to optimize towards a contextual distribution rather than a one-hot target. ICE introduces a simple yet effective optimization framework for the model to internalize new knowledge by aligning its output distributions with and without additional context. This method enhances the robustness and effectiveness of gradient-based tuning methods, preventing overfitting and preserving the model's integrity. We analyze ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, demonstrating its advantages. Experimental results confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that the integrity of the model is preserved while updating information.

The Bregman divergence, which is generated from a convex function, is commonly used as a pseudo-distance for comparing vectors or functions in continuous spaces. In contrast, defining an analog of the Bregman divergence for discrete spaces is nontrivial. Iyer & Bilmes (2012b) considered Bregman divergences on discrete domains using submodular functions as generating functions, the discrete analogs of convex functions. In this paper, we further generalize this framework to cases where the generating function is neither submodular nor supermodular, thus increasing the flexibility and representational capacity of the resulting divergence, which we term the difference-of-submodular Bregman divergence. Additionally, we introduce a learnable form of this divergence using permutation-invariant neural networks (NNs) and demonstrate through experiments that it effectively captures key structural properties in discrete data. As a result, the proposed method significantly improves the performance of existing methods on tasks such as clustering and set retrieval problems. This work addresses the challenge of defining meaningful divergences in discrete settings and provides a new tool for tasks requiring structure-preserving distance measures.

Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explanations. This behavior leads to poor generalization when the learned explanations rely on spurious correlations. In this work, we formalize $\textit{the interplay between memorization and generalization}$, showing that spurious correlations would particularly lead to poor generalization when are combined with memorization. Memorization can reduce training loss to zero, leaving no incentive to learn robust, generalizable patterns. To address this, we propose $\textit{memorization-aware training}$ (MAT), which uses held-out predictions as a signal of memorization to shift a model's logits. MAT encourages learning robust patterns invariant across distributions, improving generalization under distribution shifts.

Sparse regularization techniques are well-established in machine learning, yet their application in neural networks remains challenging due to the non-differentiability of penalties like the $L_1$ norm, which is incompatible with stochastic gradient descent. A promising alternative is shallow weight factorization, where weights are decomposed into two factors, allowing for smooth optimization of $L_1$-penalized neural networks by adding differentiable $L_2$ regularization to the factors. In this work, we introduce deep weight factorization, extending previous shallow approaches to more than two factors. We theoretically establish equivalence of our deep factorization with non-convex sparse regularization and analyze its impact on training dynamics and optimization. Due to the limitations posed by standard training practices, we propose a tailored initialization scheme and identify important learning rate requirements necessary for training factorized networks. We demonstrate the effectiveness of our deep weight factorization through experiments on various architectures and datasets, consistently outperforming its shallow counterpart and widely used pruning methods.

Estimating matrices in the symmetric positive-definite (SPD) cone is of interest for many applications ranging from computer vision to graph learning. While there exist various convex optimization-based estimators, they remain limited in expressivity due to their model-based approach. The success of deep learning motivates the use of learning-based approaches to estimate SPD matrices with neural networks in a data-driven fashion. However, designing effective neural architectures for SPD learning is challenging, particularly when the task requires additional structural constraints, such as element-wise sparsity. Current approaches either do not ensure that the output meets all desired properties or lack expressivity. In this paper, we introduce SpodNet, a novel and generic learning module that guarantees SPD outputs and supports additional structural constraints. Notably, it solves the challenging task of learning jointly SPD and sparse matrices. Our experiments illustrate the versatility and relevance of SpodNet layers for such applications.

In typical multimodal contrastive learning, such as CLIP, encoders produce one point in the latent representation space for each input. However, one-point representation has difficulty in capturing the relationship and the similarity structure of a huge amount of instances in the real world. For richer classes of the similarity, we propose the use of weighted point sets, namely, sets of pairs of weight and vector, as representations of instances. In this work, we theoretically show the benefit of our proposed method through a new understanding of the contrastive loss of CLIP, which we call symmetric InfoNCE. We clarify that the optimal similarity that minimizes symmetric InfoNCE is the pointwise mutual information, and show an upper bound of excess risk on downstream classification tasks of representations that achieve the optimal similarity. In addition, we show that our proposed similarity based on weighted point sets consistently achieves the optimal similarity. To verify the effectiveness of our proposed method, we demonstrate pretraining of text-image representation models and classification tasks on common benchmarks.

Magnetic resonance imaging (MRI) is a powerful noninvasive diagnostic imaging tool that provides unparalleled soft tissue contrast and anatomical detail. Noise contamination, especially in accelerated and/or low-field acquisitions, can significantly degrade image quality and diagnostic accuracy. Supervised learning based denoising approaches have achieved impressive performance but require high signal-to-noise ratio (SNR) labels, which are often unavailable. Self-supervised learning holds promise to address the label scarcity issue, but existing self-supervised denoising methods tend to oversmooth fine spatial features and often yield inferior performance than supervised methods. We introduce Corruption2Self (C2S), a novel score-based self-supervised framework for MRI denoising. At the core of C2S is a generalized denoising score matching (GDSM) loss, which extends denoising score matching to work directly with noisy observations by modeling the conditional expectation of higher-SNR images given further corrupted observations. This allows the model to effectively learn denoising across multiple noise levels directly from noisy data. Additionally, we incorporate a reparameterization of noise levels to stabilize training and enhance convergence, and introduce a detail refinement extension to balance noise reduction with the preservation of fine spatial features. Moreover, C2S can be extended to multi-contrast denoising by leveraging complementary information across different MRI contrasts. We demonstrate that our method achieves state-of-the-art performance among self-supervised methods and competitive results compared to supervised counterparts across varying noise conditions and MRI contrasts on the M4Raw and fastMRI dataset. The project website is available at: https://jiachentu.github.io/Corruption2Self-Self-Supervised-Denoising/.

Multiple-instance learning (MIL) was initially proposed to identify key instances within a set (bag) of instances when only one bag-level label is provided. Current deep MIL models mostly solve multi-instance problem in feature space. Nevertheless, with the increasing complexity of data, we found this paradigm faces significant risks in representation learning stage, which could lead to algorithm degradation in deep MIL models. We speculate that the degradation issue stems from the persistent drift of instances in feature space during learning. In this paper, we propose a novel Probability-Space MIL network (PSMIL) as a countermeasure. In PSMIL, a self-training alignment strategy is introduced in probability space to cope with the drift problem in feature space, and the alignment target objective is proven mathematically optimal. Furthermore, we reveal that the widely-used attention-based pooling mechanism in current deep MIL models is easily affected by the perturbation in feature space and further introduce an alternative called probability-space attention pooling. It effectively captures the key instance in each bag from feature space to probability space, and further eliminates the impact of selection drift in the pooling stage. To summarize, PSMIL seeks to solve a MIL problem in probability space rather than feature space. Experimental results illustrate that PSMIL could potentially achieve performance close to supervised learning level in complex tasks (gap within 5\%), with the incremental alignment in propability space bring more than 19\% accuracy improvements for current existing mainstream models in simulated CIFAR datasets. For existing publicly available MIL benchmarks/datasets, attention in probability space also achieves competitive performance to the state-of-the-art deep MIL models. Codes are available at \url{https://github.com/LMBDA-design/PSAMIL}.

Although it is traditionally believed that lossy image compression, such as JPEG compression, has a negative impact on the performance of deep neural networks (DNNs), it is shown by recent works that well-crafted JPEG compression can actually improve the performance of deep learning (DL). Inspired by this, we propose JPEG-DL, a novel DL framework that prepends any underlying DNN architecture with a trainable JPEG compression layer. To make the quantization operation in JPEG compression trainable, a new differentiable soft quantizer is employed at the JPEG layer, and then the quantization operation and underlying DNN are jointly trained. Extensive experiments show that in comparison with the standard DL, JPEG-DL delivers significant accuracy improvements across various datasets and model architectures while enhancing robustness against adversarial attacks. Particularly, on some fine-grained image classification datasets, JPEG-DL can increase prediction accuracy by as much as 20.9%. Our code is available on https://github.com/AhmedHussKhalifa/JPEG-Inspired-DL.git.

Despite significant progress in Anomaly Detection (AD), the robustness of existing detection methods against adversarial attacks remains a challenge, compromising their reliability in critical real-world applications such as autonomous driving. This issue primarily arises from the AD setup, which assumes that training data is limited to a group of unlabeled normal samples, making the detectors vulnerable to adversarial anomaly samples during testing. Additionally, implementing adversarial training as a safeguard encounters difficulties, such as formulating an effective objective function without access to labels. An ideal objective function for adversarial training in AD should promote strong perturbations both within and between the normal and anomaly groups to maximize margin between normal and anomaly distribution. To address these issues, we first propose crafting a pseudo-anomaly group derived from normal group samples. Then, we demonstrate that adversarial training with contrastive loss could serve as an ideal objective function, as it creates both inter- and intra-group perturbations. However, we notice that spurious negative pairs compromise the conventional contrastive loss for achieving robust AD. Spurious negative pairs are those that should be mapped closely but are erroneously separated. These pairs introduce noise and misguide the direction of inter-group adversarial perturbations. To overcome the effect of spurious negative pairs, we define opposite pairs and adversarially pull them apart to strengthen inter-group perturbations. Experimental results demonstrate our superior performance in both clean and adversarial scenarios, with a 26.1% improvement in robust detection across various challenging benchmark datasets.