Junkun Yuan

Large-scale vision-language (V-L) models have demonstrated remarkable generalization capabilities for downstream tasks through prompt tuning. However, the mechanisms behind the learned text representations are unknown, limiting further generalization gains, and the limitations are more severe when faced with the prevalent class imbalances seen in web-sourced datasets. Recent advances in the neural collapse (NC) phenomenon of vision-only models suggest that the optimal representation structure is the simplex ETF, which paves the way to study representations in V-L models. In this paper, we make the first attempt to use NC for examining the representations in V-L models via prompt tuning. It is found that NC optimality of text-to-image representations shows a positive correlation with downstream generalizability, which is more severe under class imbalance settings. To improve the representations, we propose Neural-collapse-anchored Prompt Tuning (NPT), a novel method that learns prompts with text and image representations that satisfy the same simplex Equiangular Tight Frame (ETF). NPT incorporates two regularization terms: language-modality collapse and multi-modality isomorphism; and it is compatible with other prompt tuning methods. Extensive experiments show that NPT can consistently help to improve existing prompt tuning techniques across 11 datasets for both balanced and imbalanced settings.

Universal domain adaptation (UniDA) aims to transfer knowledge from the source domain to the target domain without any prior knowledge about the label set. The challenge lies in how to determine whether the target samples belong to common categories. The mainstream methods make judgments based on the sample features, which overemphasizes global information while ignoring the most crucial local objects in the image, resulting in limited accuracy. To address this issue, we propose a Universal Attention Matching (UniAM) framework by exploiting the self-attention mechanism in vision transformer to capture the crucial object information. The proposed framework introduces a novel Compressive Attention Matching (CAM) approach to explore the core information by compressively representing attentions. Furthermore, CAM incorporates a residual-based measurement to determine the sample commonness. By utilizing the measurement, UniAM achieves domain-wise and category-wise Common Feature Alignment (CFA) and Target Class Separation (TCS). Notably, UniAM is the first method utilizing the attention in vision transformer directly to perform classification tasks. Extensive experiments show that UniAM outperforms the current state-of-the-art methods on various benchmark datasets.

Domain generalization (DG) is a prevalent problem in real-world applications, which aims to train well-generalized models for unseen target domains by utilizing several source domains. Since domain labels, i.e., which domain each data point is sampled from, naturally exist, most DG algorithms treat them as a kind of supervision information to improve the generalization performance. However, the original domain labels may not be the optimal supervision signal due to the lack of domain heterogeneity, i.e., the diversity among domains. For example, a sample in one domain may be closer to another domain, its original label thus can be the noise to disturb the generalization learning. Although some methods try to solve it by re-dividing domains and applying the newly generated dividing pattern, the pattern they choose may not be the most heterogeneous due to the lack of the metric for heterogeneity. In this paper, we point out that domain heterogeneity mainly lies in variant features under the invariant learning framework. With contrastive learning, we propose a learning potential-guided metric for domain heterogeneity by promoting learning variant features. Then we notice the differences between seeking variance-based heterogeneity and training invariance-based generalizable model. We thus propose a novel method called H eterogeneity-based Two-stage Contrastive Learning (HTCL) for the DG task. In the first stage, we generate the most heterogeneous dividing pattern with our contrastive metric. In the second stage, we employ an invariance-aimed contrastive learning by re-building pairs with the stable relation hinted by domains and classes, which better utilizes generated domain labels for generalization learning. Extensive experiments show HTCL better digs heterogeneity and yields great generalization performance.

Domain generalization (DG) aims to learn from multiple source domains a model that can generalize well on unseen target domains. Existing DG methods mainly learn the representations with invariant marginal distribution of the input features, however, the invariance of the conditional distribution of the labels given the input features is more essential for unknown domain prediction. Meanwhile, the existing of unobserved confounders which affect the input features and labels simultaneously cause spurious correlation and hinder the learning of the invariant relationship contained in the conditional distribution. Interestingly, with a causal view on the data generating process, we find that the input features of one domain are valid instrumental variables for other domains. Inspired by this finding, we propose an instrumental variable-driven DG method (IV-DG) by removing the bias of the unobserved confounders with two-stage learning. In the first stage, it learns the conditional distribution of the input features of one domain given input features of another domain. In the second stage, it estimates the relationship by predicting labels with the learned conditional distribution. Theoretical analyses and simulation experiments show that it accurately captures the invariant relationship. Extensive experiments on real-world datasets demonstrate that IV-DG method yields state-of-the-art results.

Domain generalization (DG) aims to learn from multiple known source domains a model that can generalize well to unknown target domains. The existing DG methods usually exploit the fusion of shared multi-source data to train a generalizable model. However, tremendous data is distributed across lots of places nowadays that can not be shared due to privacy policies. In this paper, we tackle the problem of federated domain generalization where the source datasets can only be accessed and learned locally for privacy protection. We propose a novel framework called Collaborative Semantic Aggregation and Calibration (CSAC) to enable this challenging problem. To fully absorb multi-source semantic information while avoiding unsafe data fusion, we conduct data-free semantic aggregation by fusing the models trained on the separated domains layer-by-layer. To address the semantic dislocation problem caused by domain shift, we further design cross-layer semantic calibration with an attention mechanism to align each semantic level and enhance domain invariance. We unify multi-source semantic learning and alignment in a collaborative way by repeating the semantic aggregation and calibration alternately, keeping each dataset localized, and the data privacy is carefully protected. Extensive experiments show the significant performance of our method in addressing this challenging problem.

Domain generalization (DG) aims to generalize the knowledge learned from multiple source domains to unseen target domains. Existing DG techniques can be subsumed under two broad categories, i.e., domain-invariant representation learning and domain manipulation. Nevertheless, it is extremely difficult to explicitly augment or generate the unseen target data. And when source domain variety increases, developing a domain-invariant model by simply aligning more domain-specific information becomes more challenging. In this paper, we propose a simple yet effective method for domain generalization, named Knowledge Distillation based Domain-invariant Representation Learning (KDDRL), that learns domain-invariant representation while encouraging the model to maintain domain-specific features, which recently turned out to be effective for domain generalization. To this end, our method incorporates multiple auxiliary student models and one student leader model to perform a two-stage distillation. In the first-stage distillation, each domain-specific auxiliary student treats the ensemble of other auxiliary students' predictions as a target, which helps to excavate the domain-invariant representation. Also, we present an error removal module to prevent the transfer of faulty information by eliminating incorrect predictions compared to the true labels. In the second-stage distillation, the student leader model with domain-specific features combines the domain-invariant representation learned from the group of auxiliary students to make the final prediction. Extensive experiments and in-depth analysis on popular DG benchmark datasets demonstrate that our KDDRL significantly outperforms the current state-of-the-art methods.

Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of 4.03% on new classes and 3.19% on harmonic-mean over eleven classification benchmarks.