Vision-Language Models (VLMs) demonstrate remarkable zero-shot generalization to unseen tasks, but fall short of the performance of supervised methods in generalizing to downstream tasks with limited data. Prompt learning is emerging as a parameter-efficient method for adapting VLMs, but state-of-the-art approaches require annotated samples. In this paper we propose a novel approach to prompt learning based on unsupervised knowledge distillation from more powerful models. Our approach, which we call Knowledge Distillation Prompt Learning (KDPL), can be integrated into existing prompt learning techniques and eliminates the need for labeled examples during adaptation. Our experiments on more than ten standard benchmark datasets demonstrate that KDPL is very effective at improving generalization of learned prompts for zero-shot domain generalization, zero-shot cross-dataset generalization, and zero-shot base-to-novel class generalization problems. KDPL requires no ground-truth labels for adaptation, and moreover we show that even in the absence of any knowledge of training class names it can be used to effectively transfer knowledge.
论文检索
输入标题、作者或关键词，从 801 篇学术成果中精准定位
Knowledge Distillation (KD) aims to learn a compact student network using knowledge from a large pre-trained teacher network, where both networks are trained on data from the same distribution. However, in practical applications, the student network may be required to perform in a new scenario (\emph{i.e.}, the target domain), which usually exhibits significant differences from the known scenario of the teacher network (\emph{i.e.}, the source domain). The traditional domain adaptation techniques can be integrated with KD in a two-stage process to bridge the domain gap, but the ultimate reliability of two-stage approaches tends to be limited due to the high computational consumption and the additional errors accumulated from both stages. To solve this problem, we propose a new one-stage method dubbed ``Direct Distillation between Different Domains" (4Ds). We first design a learnable adapter based on the Fourier transform to separate the domain-invariant knowledge from the domain-specific knowledge. Then, we build a fusion-activation mechanism to transfer the valuable domain-invariant knowledge to the student network, while simultaneously encouraging the adapter within the teacher network to learn the domain-specific knowledge of the target data. As a result, the teacher network can effectively transfer categorical knowledge that aligns with the target domain of the student network. Intensive experiments on various benchmark datasets demonstrate that our proposed 4Ds method successfully produces reliable student networks and outperforms state-of-the-art approaches.
Prompt-tuning methods for Continual Learning (CL) freeze a large pre-trained model and focus training on a few parameter vectors termed prompts. Most of these methods organize these vectors in a pool of key-value pairs, and use the input image as query to retrieve the prompts (values). However, as keys are learned while tasks progress, the prompting selection strategy is itself subject to catastrophic forgetting, an issue often overlooked by existing approaches. For instance, prompts introduced to accommodate new tasks might end up interfering with previously learned prompts. To make the selection strategy more stable, we ask a foundational model (CLIP) to select our prompt within a two-level adaptation mechanism. Specifically, the first level leverages standard textual prompts for the CLIP textual encoder, leading to stable class prototypes. The second level, instead, uses these prototypes along with the query image as keys to index a second pool. The retrieved prompts serve to adapt a pre-trained ViT, granting plasticity. In doing so, we also propose a novel residual mechanism to transfer CLIP semantics to the ViT layers. Through extensive analysis on established CL benchmarks, we show that our method significantly outperforms both state-of-the-art CL approaches and the zero-shot CLIP test. Notably, our findings hold true even for datasets with a substantial domain gap w.r.t. the pre-training knowledge of the backbone model, as showcased by experiments on satellite imagery and medical datasets. The codebase is available in the suppl. materials.
Adapt2Reward: Adapting Video-Language Models to Generalizable Robotic Rewards via Failure Prompts
PDF ↗For a general-purpose robot to operate in reality, executing a broad range of instructions across various environments is imperative. Central to the reinforcement learning and planning for such robotic agents is a generalizable reward function. Recent advances in vision-language models, such as CLIP, have shown remarkable performance in the domain of deep learning, paving the way for open-domain visual recognition. However, collecting data on robots executing various language instructions across multiple environments remains a challenge. This paper aims to transfer video-language models with robust generalization into a generalizable language-conditioned reward function, only utilizing robot video data from a minimal amount of tasks in a singular environment. Unlike common robotic datasets to train reward functions, human video-language datasets seldom include trivial failure videos. To enhance the model's ability to discriminate between successful and failed robot executions, we cluster failure data with the aspiration that the model identifies patterns in failure videos. For each cluster, we incorporate a newly trained failure prompt into the text encoder to bolster its performance in distinguishing failure from success in robot task executions. Our language-conditioned reward function shows outstanding generalization to new environments and new instructions for robot planning and reinforcement learning.
Source free domain adaptation (SFDA) aims to transfer the model trained on labeled source domain to unlabeled target domain without accessing source data. Recent SFDA methods predominantly rely on self-training, which supervise the model with pseudo labels generated from individual data samples. However, they often ignore the crucial data structure information and sample relationships that are beneficial for adaptive training. In this paper, we propose a novel hierarchical relation distillation framework, establishing multi-level relations across samples in an unsupervised manner, which fully exploits inherent data structure to guide the sample training instead of using the isolated pseudo labels. We first distinguish the source-like samples based on prediction reliability in the training process, followed by an effort on distilling knowledge to those target-specific ones by transferring both local clustering relation and global semantic relation. Specifically, we leverage the affinity with nearest neighborhood samples for local relation and consider the similarity to category-wise Gaussian Mixtures for global relation, offering complementary supervision to facilitate student learning. To validate the effectiveness of our approach, we conduct extensive experiments on four diverse benchmarks, achieving better performance compared to previous methods.
Parameter-efficient transfer learning (PETL) has emerged as a flourishing research field for its ability to adapt large pre-trained models to downstream tasks, greatly reducing trainable parameters while grappling with memory challenges during fine-tuning. To further alleviate memory burden, memory-efficient series (METL) avoid backpropagating gradients through the large backbone. However, they compromise by exclusively relying on frozen intermediate outputs and limiting the exhaustive exploration of prior knowledge from pre-trained models. Moreover, the dependency and redundancy between cross-layer features are frequently overlooked, thereby submerging more discriminative representations and causing an inherent performance gap (vs. conventional PETL methods). Hence, we propose an innovative strategy called SHERL, which stands for Synthesizing High-accuracy and Efficient-memory under Resource-Limited scenarios. Concretely, SHERL introduces a Multi-Tiered Sensing Adapter (MTSA) to decouple the entire adaptation into two successive and complementary processes. In the early route, intermediate outputs are consolidated via an anti-redundancy operation, enhancing their compatibility for subsequent interactions; thereby in the late route, utilizing minimal late pre-trained layers could alleviate the peak demand on memory overhead and regulate these fairly flexible features into more adaptive and powerful representations for new domains. We validate SHERL on 18 datasets across various vision-and-language and language-only NLP tasks. Extensive ablations demonstrate that SHERL combines the strengths of both parameter and memory-efficient techniques, performing on-par or better across diverse architectures with lower memory during fine-tuning.
Unsupervised domain adaptation (UDA) in 3D segmentation tasks presents a formidable challenge, primarily stemming from the sparse and unordered nature of point clouds. Especially for LiDAR point clouds, the domain discrepancy becomes obvious across varying capture scenes, fluctuating weather conditions, and the diverse array of LiDAR devices in use. Inspired by the remarkable generalization capabilities exhibited by the vision foundation model, SAM, in the realm of image segmentation, our approach leverages the wealth of general knowledge embedded within SAM to unify feature representations across diverse 3D domains and further solves the 3D domain adaptation problem. Specifically, we harness the corresponding images associated with point clouds to facilitate knowledge transfer and propose an innovative hybrid feature augmentation methodology, which enhances the alignment between the 3D feature space and SAM's feature space, operating at both the scene and instance levels. Our method is evaluated on many widely-recognized datasets and achieves state-of-the-art performance.
Learn from the Learnt: Source-Free Active Domain Adaptation via Contrastive Sampling and Visual Persistence
PDF ↗Domain Adaptation (DA) facilitates knowledge transfer from a source domain to a related target domain. In this paper, we investigate a practical DA paradigm, namely Source data-Free Active Domain Adaptation (SFADA), where source data becomes inaccessible during adaptation, and a minimum amount of annotation budget is available in the target domain. Without referencing the source data, new challenge emerges in identifying the most informative target samples for labeling, establishing cross-domain alignment during adaptation, and ensuring continuous performance improvements through the iterative query-and-adaptation process. We present learn from the learnt (LFTL), a novel paradigm for SFADA to leverage the learnt knowledge from the source pretrained model and actively iterated models without extra overhead. We propose Contrastive Active Sampling to learn from the hypotheses of the preceding model, thereby querying target samples that are both informative to the current model and persistently challenging throughout active learning. During adaptation, we learn from features of actively selected anchors obtained from previous intermediate models, so that the Visual Persistence-guided Adaptation can facilitate feature distribution alignment and active sample exploitation. Extensive experiments on three widely-used benchmarks show that our LFTL achieves state-of-the-art performance, superior computational efficiency and continuous improvements as the annotation budget increases. Our code will be available at https://github.com/xxx.
Continual Learning for Remote Physiological Measurement: Minimize Forgetting and Simplify Inference
PDF ↗Remote photoplethysmography (rPPG) has gained increasing attention in recent years for its ability to extract physiological signals from facial videos. Existing rPPG measurement methods have shown satisfactory performance in the intra-dataset and cross-dataset scenarios. However, they often neglect the incremental learning scenario, in which training data is presented sequentially, resulting in the issue of catastrophic forgetting. Meanwhile, mainstream class incremental learning algorithms suffer performance degradation or even fail to effectively transfer to rPPG measurement. In this paper, we present a novel and practical method to tackle continual learning for rPPG measurement. We firstly employ adapter finetuning to adapt to new tasks efficiently while enhancing the model's stability. To alleviate catastrophic forgetting without storing previous samples, we design a prototype-based augmentation strategy to reproduce the domain factors of previous tasks. Additionally, drawing inspiration from humans' problem-solving manner, an inference simplification strategy is devised to convert the potentially forgotten tasks into familiar ones for the model. To evaluate our method and enable fair comparisons, we create the first continual learning protocol for rPPG measurement. Extensive experiments demonstrate that our approach significantly surpasses the state-of-the-art methods.
Unsupervised domain adaptation (UDA) is a critical challenge in machine learning, aiming to transfer knowledge from a labeled source domain to an unlabeled target domain. In this work, we aim to improve target set accuracy in any existing UDA method by introducing an approach that utilizes pseudo-candidate sets for labeling the target data. These pseudo-candidate sets serve as a proxy for the true labels in the absence of direct supervision. To enhance the accuracy of the target domain, we propose Unsupervised Domain Adaptation refinement using Pseudo-Candidate Sets (UDPCS), a method which effectively learns to disambiguate among classes in the pseudo-candidate set. Our approach is characterized by two distinct loss functions: one that acts on the pseudo-candidate set to refine its predictions and another that operates on the labels outside the pseudo-candidate set. We use a threshold-based strategy to further guide the learning process toward accurate label disambiguation. We validate our novel yet simple approach through extensive experiments on three well-known benchmark datasets: Office-Home, VisDA, and DomainNet. Our experimental results demonstrate the efficacy of our method in achieving consistent gains on target accuracies across these datasets.
Motivated by the Parameter-Efficient Fine-Tuning (PEFT) in large language models, we propose LoRAT, a method that unveils the power of larger Vision Transformers (ViT) for tracking within laboratory-level resources. The essence of our work lies in adapting LoRA, a technique that fine-tunes a small subset of model parameters without adding inference latency, to the domain of visual tracking. However, unique challenges and potential domain gaps make this transfer not as easy as the first intuition. Firstly, a transformer-based tracker constructs unshared position embedding for template and search image. This poses a challenge for the transfer of LoRA, usually requiring consistency in the design when applied to the pre-trained backbone, to downstream tasks. Secondly, the inductive bias inherent in convolutional heads diminishes the effectiveness of parameter-efficient fine-tuning in tracking models. To overcome these limitations, we first decouple the position embeddings in transformer-based trackers into shared spatial ones and independent type ones. The shared embeddings, which describe the absolute coordinates of multi-resolution images (namely, the template and search images), are inherited from the pre-trained backbones. In contrast, the independent embeddings indicate the sources of each token and are learned from scratch. Furthermore, we design an anchor-free head solely based on a multilayer perceptron (MLP) to adapt PETR, enabling better performance with less computational overhead. With our design, 1) it becomes practical to train trackers with the ViT-g backbone on GPUs with only memory of 25.8GB (batch size of 16); 2) we reduce the training time of the L-224 variant from 35.0 to 10.8 GPU hours; 3) we improve the LaSOT SUC score from 0.703 to 0.743 with the L-224 variant; 4) we fast the inference speed of the L-224 variant from 52 to 119 FPS. Code and models will be released.
Select and Distill: Selective Dual-Teacher Knowledge Transfer for Continual Learning on Vision-Language Models
PDF ↗Large-scale vision-language models (VLMs) have shown a strong zero-shot generalization capability on unseen-domain data. However, when adapting pre-trained VLMs to a sequence of downstream tasks, they are prone to forgetting previously learned knowledge and degrade their zero-shot classification capability. To tackle this problem, we propose a unique Selective Dual-Teacher Knowledge Transfer framework that leverages the most recent fine-tuned and the original pre-trained VLMs as dual teachers to preserve the previously learned knowledge and zero-shot capabilities, respectively. With only access to an unlabeled reference dataset, our proposed framework performs a selective knowledge distillation mechanism by measuring the feature discrepancy from the dual teacher VLMs. Consequently, our selective dual-teacher knowledge distillation would mitigate catastrophic forgetting of previously learned knowledge while preserving the zero-shot capabilities from pre-trained VLMs. Through extensive experiments on benchmark datasets, we show that our proposed framework is favorable against state-of-the-art continual learning approaches for preventing catastrophic forgetting and zero-shot degradation.
Semi-supervised graph domain adaptation, as a subfield of graph transfer learning, seeks to precisely annotate unlabeled target graph nodes by leveraging transferable features acquired from the limited labeled source nodes. However, most existing studies often directly utilize graph convolutional networks (GCNs)-based feature extractors to capture domain-invariant node features, while neglecting the issue that GCNs are insufficient in collecting complex structure information in graph. Considering the importance of graph structure information in encoding the complex relationship among nodes and edges, this paper aims to utilize such powerful information to assist graph transfer learning. To achieve this goal, we develop a novel framework called HOGDA. Concretely, HOGDA introduces a high-order structure information mixing (HSIM) module to effectively capture abundant structure information in graph, greatly enhancing the feature extractor's ability to adapt across different domains. Moreover, to achieve fine-grained feature distributions alignment, a novel strategy called adaptive weighted domain alignment (AWDA) is proposed to dynamically adjust the node weight during adversarial domain adaptation process, effectively boosting the model's transfer ability. Furthermore, to mitigate the overfitting phenomenon caused by limited source labeled nodes, we also design a trust-aware node clustering (TNC) strategy to guide the unlabeled nodes to achieve discriminative clustering. Extensive experimental results show that our HOGDA outperforms the state-of-the-art methods on various transfer tasks.
Video Tube Retrieval (VTR) has attracted wide attention in the multi-modal domain, aiming to accurately localize the spatial-temporal tube in videos based on the natural language description. Despite the remarkable progress, existing VTR models trained on a specific domain (source domain) often perform unsatisfactory in another domain (target domain), due to the domain gap. Toward this issue, we introduce the learning strategy, Unsupervised Domain Adaptation, into the VTR task (UDA-VTR), which enables the knowledge transfer from the labeled source domain to the unlabeled target domain without additional manual annotations. An intuitive solution is generating the pseudo labels for the target domain samples with the fully trained source model and fine-tuning the source model on the target domain with pseudo labels. However, the existing domain gap gives rise to two problems for this process: (1) The transfer of model parameters across domains may introduce source domain bias into target domain features, significantly impacting the feature-based prediction for target domain samples. (2) The pseudo labels tend to identify video tubes that are widely present in the source domain, rather than accurately localizing the correct video tubes specific to the target domain samples. To address the above issues, we propose the unsupervised domain adaptation model via Hierarchical dEbiAsing and noisy corRecTion (HEART) for cross-domain video tube retrieval, which contains two characteristic modules: Layered Feature Debiasing (including the adversarial feature alignment and the graph based alignment) and Pseudo Label Refinement. Extensive experiments prove the effectiveness of our HEART model by significantly surpassing the state-of-the-arts.
Black-box domain adaptation treats the source domain model as a black box. During the transfer process, the only available information about the target domain is the noisy labels output by the black-box model. This poses significant challenges for domain adaptation. Conventional approaches typically tackle the black-box noisy label problem from two aspects: self-knowledge distillation and pseudo-label denoising, both achieving limited performance due to limited knowledge information. To mitigate this issue, we explore the potential of off-the-shelf vision-language (ViL) multimodal models with rich semantic information for black-box domain adaptation by introducing an Adversarial Experts Model (AEM). Specifically, our target domain model is designed as one feature extractor and two classifiers, trained over two stages: In the knowledge transferring stage, with a shared feature extractor, the black-box source model and the ViL model act as two distinct experts for joint knowledge contribution, guiding the learning of one classifier each. While contributing their respective knowledge, the experts are also updated due to their own limitation and bias. In the adversarial alignment stage, to further distill expert knowledge to the target domain model, adversarial learning is conducted between the feature extractor and the two classifiers. A new consistency-max loss function is proposed to measure two classifier consistency and further improve classifier prediction certainty. Extensive experiments on multiple datasets demonstrate the effectiveness of our approach. Code is available at https://github.com/singinger/AEM.
Continual learning emerges as a framework that trains the model on a sequence of tasks without forgetting previously learned knowledge, which has been applied in multiple multimodal scenarios. Recently, prompt-based continual learning has achieved excellent domain adaptability and knowledge transfer through prompt generation. However, existing methods mainly focus on designing the architecture of a generator, neglecting the importance of providing effective guidance for training the generator. To address this issue, we propose Generating Prompts in Latent Space (GPLS), which considers prompts as latent variables to account for the uncertainty of prompt generation and aligns with the fact that prompts are inserted into the hidden layer outputs and exert an implicit influence on classification. GPLS adopts a trainable encoder to encode task and feature information into prompts with reparameterization technique, and provides refined and targeted guidance for the training process through the evidence lower bound (ELBO) related to Mahalanobis distance. Extensive experiments demonstrate that GPLS achieves state-of-the-art performance on various benchmarks. Our code is available at https://github.com/Hifipsysta/GPLS.
Large-scale pretrained image-language models have shown remarkable performance recently. However, building a video-language model is more challenging due to the complexity of video and the difficulty of collecting high-quality data. This paper builds a video-language model in an adaptive manner, which transfers the knowledge from the image domain and can achieve state-of-the-art performance without any further massive video pretraining. The main contributions include a Visual Perception Adapter that seamlessly and efficiently adapts a pretrained image-language model to the video domain and a fine-grained contrastive learning with Inter-modal Token Alignment that bridges semantic gaps between vision, audio, and language with less data. The proposed model is evaluated on video captioning and retrieval. Experiments demonstrate that the proposed model exhibits competitive performance compared to models pretrained on millions of video-text pairs. Notably, our model's CIDEr and R@1 scores on the MSR-VTT dataset exceed the existing state-of-the-art by 6.3% and 1.3%.
The advent of High Dynamic Range/Wide Color Gamut (HDR/WCG) display technology has made significant progress in providing exceptional richness and vibrancy for the human visual experience. However, the widespread adoption of HDR/WCG images is hindered by their substantial storage requirements, imposing significant bandwidth challenges during distribution. Besides, HDR/WCG images are often tone-mapped into Standard Dynamic Range (SDR) versions for compatibility, necessitating the usage of inverse Tone Mapping (iTM) techniques to reconstruct their original representation. In this work, we propose a meta-transfer learning framework for practical HDR/WCG media transmission by embedding image-wise metadata into their SDR counterparts for later iTM reconstruction. Specifically, we devise a meta-learning strategy to pre-train a lightweight multilayer perceptron (MLP) model that maps SDR pixels to HDR/WCG ones on an external dataset, resulting in a domain-wise iTM model. Subsequently, for the transfer learning process of each HDR/WCG image, we present a spatial-aware online mining mechanism to select challenging training pairs to adapt the meta-trained model to an image-wise iTM model. Finally, the adapted MLP, embedded as metadata, is transmitted alongside the SDR image, facilitating the reconstruction of the original image on HDR/WCG displays. We conduct extensive experiments and evaluate the proposed framework with diverse metrics. Compared with existing solutions, our framework shows superior performance in fidelity, minimal latency, and negligible overhead. The codes are available at https://github.com/pjliu3/MLP_iTM.
Incremental Object Detection (IOD) simulates the dynamic data flow in real-world applications, which require detectors to learn new classes or adapt to new domains while retaining knowledge from previous tasks. Most existing IOD methods focus only on class incremental learning, assuming all data comes from the same domain. However, this is hardly achievable in practical applications, as images collected under different conditions often exhibit completely different characteristics, such as lighting, weather, style, etc. Class IOD methods suffer from performance degradation in these scenarios with domain shifts. To bridge domain shifts and category gaps in IOD, we propose Purified Distillation (PD), where we use a set of trainable queries to transfer the teacher's attention on old tasks to the student and adopt the gradient reversal layer to guide the student to learn the teacher's feature space structure from a micro perspective, which has not been extensively studied in previous works. Meanwhile, PD combines classification confidence with localization confidence to purify the most meaningful output nodes, so that the student model inherits a more comprehensive teacher knowledge. Extensive experiments across various IOD settings on six widely used datasets show that PD significantly outperforms state-of-the-art methods. Even after five steps of incremental learning, our method can preserve 60.6% mAP on the first task, while compared methods can only maintain up to 55.9%.
Scene segmentation via unsupervised domain adaptation (UDA) enables the transfer of knowledge acquired from source synthetic data to real-world target data, which largely reduces the need for manual pixel-level annotations in the target domain. To facilitate domain-invariant feature learning, existing methods typically mix data from both the source domain and target domain by simply copying and pasting pixels. Such vanilla methods are usually sub-optimal since they do not take into account how well the mixed layouts correspond to real-world scenarios. Real-world scenarios are with an inherent layout. The model suffers from confusion in predicting the target domain due to the unrealistic mixing. For instance, it is not reasonable to directly paste the near "pedestrian'' pixels into the remote "sky'' area. Based on such observation, we propose a depth-aware framework to explicitly leverage depth estimation to mix categories and facilitate two complementary tasks, i.e., segmentation and depth learning in an end-to-end manner. Besides, several public datasets do not provide depth annotation. Therefore, we leverage the off-the-shelf depth estimation network to obtain the pseudo depth. Extensive experiments show that our methods, even with pseudo depth, achieve competitive performance, i.e., 77.7 mIoU on GTA → Cityscapes and 69.3 mIoU on Synthia → Cityscapes.