Contrastive Language Image Pre-training (CLIP) has recently demonstrated success across various tasks due to superior feature representation empowered by image-text contrastive learning. However, the instance discrimination method used by CLIP can hardly encode the semantic structure of training data. To handle this limitation, cluster discrimination has been proposed through iterative cluster assignment and classification. Nevertheless, most cluster discrimination approaches only define a single pseudo-label for each image, neglecting multi-label signals in the image. In this paper, we propose a novel Multi-Label Cluster Discrimination method named MLCD to enhance representation learning. In the clustering step, we first cluster the large-scale LAION-400M dataset into one million centers based on off-the-shelf embedding features. Considering that natural images frequently contain multiple visual objects or attributes, we select the multiple closest centers as auxiliary class labels. In the discrimination step, we design a novel multi-label classification loss, which elegantly separates losses from positive classes and negative classes, and alleviates ambiguity on decision boundary. We validate the proposed multi-label cluster discrimination method with experiments on different scales of models and pre-training datasets. Experimental results show that our method achieves state-of-the-art performance on multiple downstream tasks including linear probe, zero-shot classification, and image-text retrieval.
论文检索
输入标题、作者或关键词，从 7,876 篇学术成果中精准定位
Popular representation learning methods encourage feature invariance under transformations applied at the input. However, in 3D perception tasks like object localization and segmentation, outputs are naturally equivariant to some transformations, such as rotation. Using pre-training loss functions that encourage equivariance of features under certain transformations provides a strong self-supervision signal while also retaining information of geometric relationships between transformed feature representations. This can enable improved performance in downstream tasks that are equivariant to such transformations. In this paper, we propose a spatio-temporal equivariant learning framework by considering both spatial and temporal augmentations jointly. Our experiments show that the best performance arises with a pre-training approach that encourages equivariance to translation, scaling, and flip, rotation and scene flow. For spatial augmentations, we find that depending on the transformation, either a contrastive objective or an equivariance-by-classification objective yields best results. To leverage real-world object deformations and motion, we consider sequential LiDAR scene pairs and develop a novel 3D scene flow-based equivariance objective that leads to improved performance overall. We show our pre-training method for 3D object detection which outperforms existing equivariant and invariant approaches in many settings.
Concept Bottleneck Models (CBMs) ground image classification on human-understandable concepts to allow for interpretable model decisions. Crucially, the CBM design inherently allows for human interventions, in which expert users are given the ability to modify potentially misaligned concept choices to influence the decision behavior of the model in an interpretable fashion. However, existing approaches often require numerous human interventions per image to achieve strong performances, posing practical challenges in scenarios where obtaining human feedback is expensive. In this paper, we find that this is noticeably driven by an independent treatment of concepts during intervention, wherein a change of one concept does not influence the use of other ones in the model's final decision. To address this issue, we introduce a trainable concept intervention realignment module, which leverages concept relations to realign concept assignments post-intervention. Across standard, real-world benchmarks, we find that concept re-alignment can significantly improve intervention efficacy; significantly reducing the number of interventions needed to reach a target classification performance or concept prediction accuracy. In addition, it easily integrates into existing concept-based architectures without requiring changes to the models themselves. This reduced cost of human-model collaboration is crucial to enhance the feasibility of CBMs in resource-constrained environments.
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.
The unprecedented advancements in Large Language Models (LLMs) have shown a profound impact on natural language processing but are yet to fully embrace the realm of 3D understanding. This paper introduces PointLLM, a preliminary effort to fill this gap, empowering LLMs to understand point clouds and offering a new avenue beyond 2D data. PointLLM understands colored object point clouds with human instructions and generates contextually appropriate responses, illustrating its grasp of point clouds and common sense. Specifically, it leverages a point cloud encoder with a powerful LLM to effectively fuse geometric, appearance, and linguistic information. To overcome the scarcity of point-text instruction following data, we developed an automated data generation pipeline, collecting a large-scale dataset of more than 730K samples with 660K different objects, which facilitates the adoption of the two-stage training strategy prevalent in MLLM development. Additionally, we address the absence of appropriate benchmarks and the limitations of current evaluation metrics by proposing two novel benchmarks: Generative 3D Object Classification and 3D Object Captioning, which are supported by new, comprehensive evaluation metrics derived from human and GPT analyses. Through exploring various training strategies, we develop PointLLM, significantly surpassing 2D and 3D baselines, with a notable achievement in human-evaluated object captioning tasks where it surpasses human annotators in over 50% of the samples. Codes, datasets, and benchmarks are available at https://github.com/OpenRobotLab/PointLLM.
This paper addresses the critical issue of miscalibration in CLIP-based adaptation models in the challenging scenario of out-of-distribution samples, which has been overlooked in the existing literature on CLIP adaptation. We empirically demonstrate that popular CLIP adaptation approaches, such as adapters, prompt learning, and test-time prompt tuning, substantially degrade the calibration capabilities of the zero-shot baseline in the presence of distributional drift. We identify the increase in logit ranges as the underlying cause of miscalibration of CLIP adaptation methods, contrasting with previous work on calibrating fully supervised models. Motivated by these observations, we present a simple and model-agnostic solution to mitigate miscalibration, by scaling the logit range of each sample based on its zero-shot prediction logits. We explore three different alternatives to achieve this, which can be either integrated during adaptation, or directly used at inference time. Comprehensive experiments on popular OOD classification benchmarks demonstrate the effectiveness of the proposed approaches in mitigating miscalibration while maintaining discriminative performance, whose improvements are consistent across the three families of these increasingly popular approaches.
CardiacNet: Learning to Reconstruct Abnormalities for Cardiac Disease Assessment from Echocardiogram Videos
PDF ↗Echocardiography plays a crucial role in analyzing cardiac function and diagnosing cardiac diseases. Current deep neural network methods primarily aim to enhance diagnosis accuracy by incorporating prior knowledge, such as segmenting cardiac structures or lesions annotated by human experts. However, diagnosing the inconsistent behaviours of the heart, which exist across both spatial and temporal dimensions, remains extremely challenging. For instance, the analysis of cardiac motion acquires both spatial and temporal information from the heartbeat cycle. To address this issue, we propose a novel reconstruction-based approach named CardiacNet to learn a better representation of local cardiac structures and motion abnormalities through echocardiogram videos. CardiacNet is accompanied by the Consistency Deformation Codebook (CDC) and the Consistency Deformed-Discriminator (CDD) to learn the commonalities across abnormal and normal samples by incorporating cardiac prior knowledge. In addition, we propose benchmark datasets named CardiacNet-PAH and CardiacNet-ASD for evaluating the effectiveness of cardiac disease assessment. In experiments, our CardiacNet can achieve state-of-the-art results in downstream tasks, i.e., classification and regression, on public datasets CAMUS, EchoNet, and our datasets. The codes and datasets will be publicly available.
Subsampling layers play a crucial role in deep nets by discarding a portion of an activation map to reduce its spatial dimensions. This encourages the deep net to learn higher-level representations. Contrary to this motivation, we hypothesize that the discarded activations are useful and can be incorporated on the fly to improve models' prediction. To validate our hypothesis, we propose a search and aggregate method to find useful activation maps to be used at test-time. We applied our approach to the task of image classification and semantic segmentation. Extensive experiments over nine different architectures on ImageNet, CityScapes, and ADE20K show that our method consistently improves model test-time performance. Additionally, it complements existing test-time augmentation techniques to provide further performance gains.
Self-supervised learning (SSL) has proven effective in learning high-quality representations for various downstream tasks, with a primary focus on semantic tasks. However, its application in geometric tasks remains underexplored, partially due to the absence of a standardized evaluation method for geometric representations. To address this gap, we introduce a novel pose-estimation benchmark for assessing SSL geometric representations, which demands training without semantic or pose labels and achieving proficiency in both semantic and geometric downstream tasks. On this benchmark, we study enhancing SSL geometric representations without sacrificing semantic classification accuracy. We find that leveraging mid-layer representations improves pose-estimation performance by 10-20%. Further, we introduce an unsupervised trajectory-regularization loss, which improves performance by an additional 4% and improves generalization ability on out-of-distribution data. We hope the proposed benchmark and methods offer new insights and improvements in self-supervised geometric representation learning.
Video action detection (VAD) aims to detect actors and classify their actions in a video. We figure that VAD suffers more from classification rather than localization of actors. Hence, we analyze how prevailing methods form features for classification and find that they prioritize actor regions for classification, yet often overlooking the essential contextual information necessary for accurate classification. Accordingly, we propose to reduce the model's bias toward the actor itself and encourage it to pay attention to the context that is relevant to each action class. By assigning a class-dedicated query to each action class, the model can dynamically determine where to focus for effective classification. The proposed method demonstrates superior performance on three challenging benchmarks while using significantly fewer parameters and less computation.
Deep neural networks (DNNs) for image classification are known to be susceptible to many diversified universal adversarial perturbations (UAPs), where each UAP successfully attacks a large but substantially different set of images. Properly combining the diversified UAPs can significantly improve the attack effectiveness, as the sets of images successfully attacked by different UAPs are complementary to each other. In this paper, we study this novel type of attack by developing a cocktail universal adversarial attack framework. The key idea is to train a set of diversified UAPs and a selection neural network at the same time, such that the selection neural network can choose the most effective UAP when attacking a new target image. Due to the simplicity and effectiveness of the cocktail attack framework, it can be generally used to significantly boost the attack effectiveness of many classic single-UAP methods that use a single UAP to attack all target images. The proposed cocktail attack framework is also able to perform real-time attacks as it does not require additional training or fine-tuning when attacking new target images. Extensive experiments demonstrate the outstanding performance of cocktail attacks.
Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than 45% on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning.
Advancements in open-source pre-trained backbones make it relatively easy to fine-tune a model for new tasks. However, this lowered entry barrier poses potential risks, e.g., bad actors developing models for harmful applications. A question arises: "Is possible to develop a pre-trained model that is difficult to fine-tune for certain downstream tasks?" To begin studying this, we focus on few-shot classification (FSC). Specifically, we investigate methods to make FSC more challenging for a set of restricted classes while maintaining the performance of other classes. We propose to meta-learn over the pre-trained backbone in a manner that renders it a ``poor initialization''. Our proposed Learning to Obstruct (LTO) algorithm successfully obstructs four FSC methods across three datasets, including ImageNet and CIFAR100 for image classification, as well as CelebA for attribute classification.
Recent studies have revealed the vulnerability of deep neural networks (DNNs) to various backdoor attacks, where the behavior of DNNs can be compromised by utilizing certain types of triggers or poisoning mechanisms. State-of-the-art (SOTA) defenses employ too-sophisticated mechanisms that require either a computationally expensive adversarial search module for reverse-engineering the trigger distribution or an over-sensitive hyper-parameter selection module. Moreover, they offer sub-par performance in challenging scenarios, e.g., limited validation data and strong attacks. In this paper, we propose--{\em Neural mask Fine-Tuning (NFT)}-- with an aim to optimally re-organize the neuron activities in a way that the effect of the backdoor is removed. {\color{black}Utilizing a simple data augmentation like MixUp, NFT relaxes the trigger synthesis process and eliminates the requirement of the adversarial search module.} Our study further reveals that direct weight fine-tuning under limited validation data results in poor post-purification clean test accuracy, primarily due to \emph{overfitting issue}. To overcome this, we propose to fine-tune neural masks instead of model weights. In addition, a \emph{mask regularizer} has been devised to further mitigate the model drift during the purification process. The distinct characteristics of NFT render it highly efficient in both runtime and sample usage, as it can remove the backdoor even when a single sample is available from each class. We validate the effectiveness of NFT through extensive experiments covering the tasks of image classification, object detection, video action recognition, 3D point cloud, and natural language processing. We evaluate our method against 14 different attacks (LIRA, WaNet, etc.) on 11 benchmark data sets (ImageNet, UCF101, Pascal VOC, ModelNet, OpenSubtitles2012, etc).
Large-scale image-text pre-trained models enable zero-shot classification and provide consistent accuracy across various data distributions. Nonetheless, optimizing these models in downstream tasks typically requires fine-tuning, which reduces generalization to out-of-distribution (OOD) data and demands extensive computational resources. We introduce Robust Adapter (R-Adapter), a novel method for fine-tuning zero-shot models to downstream tasks while simultaneously addressing both these issues. Our method integrates lightweight modules into the pre-trained model and employs novel self-ensemble techniques to boost OOD robustness and reduce storage expenses substantially. Furthermore, we propose MPM-NCE loss designed for fine-tuning on vision-language downstream tasks. It ensures precise alignment of multiple image-text pairs and discriminative feature learning. By extending the benchmark for robust fine-tuning beyond classification to include diverse tasks such as cross-modal retrieval and open vocabulary segmentation, we demonstrate the broad applicability of R-Adapter. Our extensive experiments demonstrate that R-Adapter achieves state-of-the-art performance across a diverse set of tasks, tuning only 13% of the parameters of the CLIP encoders.
Online video understanding often relies on individual frames, leading to frame-by-frame predictions. Recent advancements such as Online Temporal Action Localization (OnTAL), extend this approach to instance-level predictions. However, existing methods mainly focus on short-term context, neglecting historical information. To address this, we introduce the History-Augmented Anchor Transformer (HAT) Framework for OnTAL. By integrating historical context, our framework enhances the synergy between long-term and short-term information, improving the quality of anchor features crucial for classification and localization. We evaluate our model on both procedural egocentric (PREGO) datasets (EGTEA and EPIC) and standard non-PREGO OnTAL datasets (THUMOS and MUSES). Results show that our model outperforms state-of-the-art approaches significantly on PREGO datasets and achieves comparable or slightly superior performance on non-PREGO datasets, underscoring the importance of leveraging long-term history, especially in procedural and egocentric action scenarios. Code is available at: (Online video understanding often relies on individual frames, leading to frame-by-frame predictions. Recent advancements such as Online Temporal Action Localization (OnTAL), extend this approach to instance-level predictions. However, existing methods mainly focus on short-term context, neglecting historical information. To address this, we introduce the History-Augmented Anchor Transformer (HAT) Framework for OnTAL. By integrating historical context, our framework enhances the synergy between long-term and short-term information, improving the quality of anchor features crucial for classification and localization. We evaluate our model on both procedural egocentric (PREGO) datasets (EGTEA and EPIC) and standard non-PREGO OnTAL datasets (THUMOS and MUSES). Results show that our model outperforms state-of-the-art approaches significantly on PREGO datasets and achieves comparable or slightly superior performance on non-PREGO datasets, underscoring the importance of leveraging long-term history, especially in procedural and egocentric action scenarios. Code is available at: https://github.com/sakibreza/ECCV24-HAT/.
Existing human mesh recovery (HMR) methods commonly leverage the global or dense-annotations-based local features to produce a single prediction from input image. However, the compressed global feature and local features disrupt the spatial geometry of human body and make it hard to capture the local dynamics, resulting in visual-mesh misalignment. Moreover, dense annotations are labor-intensive and expensive. Toward above issues, we propose a global-to-pixel wise prediction framework to preserve spatial information and obtain precise visual-mesh alignments for top-down HMR. Specifically, we present an adaptive 2D Keypoint-Guided Local Encoding Module to enable per-pixel features to capture fine-grained body part information with structure and local context maintained. The acquisition of local features rely exclusively on sparse 2D keypoint guidance without any dense annotations or heuristics keypoint-based ROI pooling. The enhanced pixel features are used to predict residuals for rectifying initial estimation produced by global feature. Secondly, we introduce a Dynamic Matching Strategy that determines positive/negative pixels by only calculating the classification and 2D keypoint costs to further improve visual-mesh alignments. The comprehensive experiments demonstrate the effectiveness of network design. Our framework outperforms previous local regression methods by a large margin and achieves state-of-the-art performance on Human3.6m and 3DPW datasets.
Token compression expedites the training and inference of Vision Transformers (ViTs) by reducing the number of the redundant tokens, e.g., pruning inattentive tokens or merging similar tokens. However, when applied to downstream tasks, these approaches suffer from significant performance drop when the compression degrees are mismatched between training and inference stages, which limits the application of token compression on off-the-shelf trained models. In this paper, we propose a model arithmetic framework to decouple the compression degrees between the two stages. In advance, we additionally perform a fast parameter-efficient self-distillation stage on the pre-trained model to obtain a small plugin, called Token Compensator (ToCom), which describes the gap between models across different compression degrees. During inference, ToCom can be directly inserted into any downstream off-the-shelf models with any mismatched training and inference compression degrees to acquire universal performance improvements without further training. Experiments on over 20 downstream tasks demonstrate the effectiveness of our framework. On CIFAR100, fine-grained visual classification, and VTAB-1k benchmark, ToCom can yield up to a maximum improvement of 2.3%, 1.5%, and 2.0% in the average performance of DeiT-B, respectively.
ViC-MAE: Self-Supervised Representation Learning from Images and Video with Contrastive Masked Autoencoders
PDF ↗We propose ViC-MAE, a model that combines both Masked AutoEncoders (MAE) and contrastive learning. ViC-MAE is trained using a global feature obtained by pooling the local representations learned under an MAE reconstruction loss and leveraging this representation under a contrastive objective across images and video frames. We show that visual representations learned under ViC-MAE generalize well to video and image classification tasks. Particularly, ViC-MAE obtains state-of-the-art transfer learning performance from video to images on Imagenet-1k compared to the recently proposed OmniMAE by achieving a top-1 accuracy of 86% (+1.3% absolute improvement) when trained on the same data and 87.1% (+2.4% absolute improvement) when training on extra data. At the same time, ViC-MAE outperforms most other methods on video benchmarks by obtaining 75.9% top-1 accuracy on the challenging Something something-v2 video benchmark. When training on videos and images from diverse datasets, our method maintains a balanced transfer-learning performance between video and image classification benchmarks, coming only as a close second to the best-supervised method. Source code and model checkpoints will be released with this paper.
The activation function plays a crucial role in model optimisation, yet the optimal choice remains unclear. For example, the Sigmoid activation is the de-facto activation in balanced classification tasks, however, in imbalanced classification, it proves inappropriate due to bias towards frequent classes. In this work, we delve deeper in this phenomenon by performing a comprehensive statistical analysis in the classification and intermediate layers of both balanced and imbalanced networks and we empirically show that aligning the activation function with the data distribution, enhances the performance in both balanced and imbalanced tasks. To this end, we propose the Adaptive Parametric Activation (APA) function, a novel and versatile activation function that unifies most common activation functions under a single formula. APA can be applied in both intermediate layers and attention layers, significantly outperforming the state-of-the-art on several imbalanced benchmarks such as ImageNet-LT, iNaturalist2018, Places-LT, CIFAR100-LT and LVIS and balanced benchmarks such as ImageNet1K, COCO and V3DET. Code is provided in the supplementary material.