Duo Li

Yifan Pu, Zhuofan Xia, Jiayi Guo et al.

This paper identifies significant redundancy in the query-key interactions within self-attention mechanisms of diffusion transformer models, particularly during the early stages of denoising diffusion steps. In response to this observation, we present a novel diffusion transformer framework incorporating an additional set of mediator tokens to engage with queries and keys separately. By modulating the number of mediator tokens during the denoising generation phases, our model initiates the denoising process with a precise, non-ambiguous stage and gradually transitions to a phase enriched with detail. Concurrently, integrating mediator tokens simplifies the attention module's complexity to a linear scale, enhancing the efficiency of global attention processes. Additionally, we propose a time-step dynamic mediator token adjustment mechanism that further decreases the required computational FLOPs for generation, simultaneously facilitating the generation of high-quality images within the constraints of varied inference budgets. Extensive experiments demonstrate that the proposed method can improve the generated image quality while also reducing the inference cost of diffusion transformers. When integrated with the recent work SiT, our method achieves a state-of-the-art FID score of 2.01. The source code is available at https://github.com/LeapLabTHU/Attention-Mediators.

Guimei Cao, Zhanzhan Cheng, Yunlu Xu et al.

Expandable networks have demonstrated their advantages in dealing with catastrophic forgetting problem in incremental learning. Considering that different tasks may need different structures, recent methods design dynamic structures adapted to different tasks via sophisticated skills. Their routine is to search expandable structures first and then train on the new tasks, which, however, breaks tasks into multiple training stages, leading to suboptimal or overmuch computational cost. In this paper, we propose an end-to-end trainable adaptively expandable network named E2-AEN, which dynamically generates lightweight structures for new tasks without any accuracy drop in previous tasks. Specifically, the network contains a serial of powerful feature adapters for augmenting the previously learned representations to new tasks, and avoiding task interference. These adapters are controlled via an adaptive gate-based pruning strategy which decides whether the expanded structures can be pruned, making the network structure dynamically changeable according to the complexity of the new tasks. Moreover, we introduce a novel sparsity-activation regularization to encourage the model to learn discriminative features with limited parameters. E2-AEN reduces cost and can be built upon any feed-forward architectures in an end-to-end manner. Extensive experiments on both classification (i.e., CIFAR and VDD) and detection (i.e., COCO, VOC and ICCV2021 SSLAD challenge) benchmarks demonstrate the effectiveness of the proposed method, which achieves the new remarkable results.

Jia Hu, Zhexi Lian, Haoran Wang et al.

Adaptive Cruise Control (ACC) systems have been widely commercialized in recent years. However, existing ACC systems remain vulnerable to close-range cut-ins, a behavior that resembles "road bullying". To address this issue, this research proposes an Anti-bullying Adaptive Cruise Control (AACC) approach, which is capable of proactively protecting right-of-way against such "road bullying" cut-ins. To handle diverse "road bullying" cut-in scenarios smoothly, the proposed approach first leverages an online Inverse Optimal Control (IOC) based algorithm for individual driving style identification. Then, based on Stackelberg competition, a game-theoretic-based motion planning framework is presented in which the identified individual driving styles are utilized to formulate cut-in vehicles' reaction functions. By integrating such reaction functions into the ego vehicle's motion planning, the ego vehicle could consider cut-in vehicles' all possible reactions to find its optimal right-of-way protection maneuver. To the best of our knowledge, this research is the first to model vehicles' interaction dynamics and develop an interactive planner that adapts cut-in vehicle's various driving styles. Simulation results show that the proposed approach can prevent "road bullying" cut-ins and be adaptive to different cut-in vehicles' driving styles. It can improve safety and comfort by up to 79.8% and 20.4%. The driving efficiency has benefits by up to 19.33% in traffic flow. The proposed approach can also adopt more flexible driving strategies. Furthermore, the proposed approach can support real-time field implementation by ensuring less than 50 milliseconds computation time.

Wenhao Jiang, Duo Li, Menghan Hu et al.

To tackle the issues of catastrophic forgetting and overfitting in few-shot class-incremental learning (FSCIL), previous work has primarily concentrated on preserving the memory of old knowledge during the incremental phase. The role of pre-trained model in shaping the effectiveness of incremental learning is frequently underestimated in these studies. Therefore, to enhance the generalization ability of the pre-trained model, we propose Learning with Prior Knowledge (LwPK) by introducing nearly free prior knowledge from a few unlabeled data of subsequent incremental classes. We cluster unlabeled incremental class samples to produce pseudo-labels, then jointly train these with labeled base class samples, effectively allocating embedding space for both old and new class data. Experimental results indicate that LwPK effectively enhances the model resilience against catastrophic forgetting, with theoretical analysis based on empirical risk minimization and class distance measurement corroborating its operational principles. The source code of LwPK is publicly available at: \url{https://github.com/StevenJ308/LwPK}.

Kuo Yang, Duo Li, Menghan Hu et al.

For semi-supervised learning with imbalance classes, the long-tailed distribution of data will increase the model prediction bias toward dominant classes, undermining performance on less frequent classes. Existing methods also face challenges in ensuring the selection of sufficiently reliable pseudo-labels for model training and there is a lack of mechanisms to adjust the selection of more reliable pseudo-labels based on different training stages. To mitigate this issue, we introduce uncertainty into the modeling process for pseudo-label sampling, taking into account that the model performance on the tailed classes varies over different training stages. For example, at the early stage of model training, the limited predictive accuracy of model results in a higher rate of uncertain pseudo-labels. To counter this, we propose an Uncertainty-Aware Dynamic Threshold Selection (UDTS) approach. This approach allows the model to perceive the uncertainty of pseudo-labels at different training stages, thereby adaptively adjusting the selection thresholds for different classes. Compared to other methods such as the baseline method FixMatch, UDTS achieves an increase in accuracy of at least approximately 5.26%, 1.75%, 9.96%, and 1.28% on the natural scene image datasets CIFAR10-LT, CIFAR100-LT, STL-10-LT, and the medical image dataset TissueMNIST, respectively. The source code of UDTS is publicly available at: https://github.com/yangk/UDTS.

Xiangtai Li, Hao He, Xia Li et al.

Aerial Image Segmentation is a particular semantic segmentation problem and has several challenging characteristics that general semantic segmentation does not have. There are two critical issues: The one is an extremely foreground-background imbalanced distribution, and the other is multiple small objects along with the complex background. Such problems make the recent dense affinity context modeling perform poorly even compared with baselines due to over-introduced background context. To handle these problems, we propose a point-wise affinity propagation module based on the Feature Pyramid Network (FPN) framework, named PointFlow. Rather than dense affinity learning, a sparse affinity map is generated upon selected points between the adjacent features, which reduces the noise introduced by the background while keeping efficiency. In particular, we design a dual point matcher to select points from the salient area and object boundaries, respectively. Experimental results on three different aerial segmentation datasets suggest that the proposed method is more effective and efficient than state-of-the-art general semantic segmentation methods. Especially, our methods achieve the best speed and accuracy trade-off on three aerial benchmarks. Further experiments on three general semantic segmentation datasets prove the generality of our method. Code will be provided in (https: //github.com/lxtGH/PFSegNets).

Duo Li, Jie Hu, Changhu Wang et al.

Convolution has been the core ingredient of modern neural networks, triggering the surge of deep learning in vision. In this work, we rethink the inherent principles of standard convolution for vision tasks, specifically spatial-agnostic and channel-specific. Instead, we present a novel atomic operation for deep neural networks by inverting the aforementioned design principles of convolution, coined as involution. We additionally demystify the recent popular self-attention operator and subsume it into our involution family as an over-complicated instantiation. The proposed involution operator could be leveraged as fundamental bricks to build the new generation of neural networks for visual recognition, powering different deep learning models on several prevalent benchmarks, including ImageNet classification, COCO detection and segmentation, together with Cityscapes segmentation. Our involution-based models improve the performance of convolutional baselines using ResNet-50 by up to 1.6% top-1 accuracy, 2.5% and 2.4% bounding box AP, and 4.7% mean IoU absolutely while compressing the computational cost to 66%, 65%, 72%, and 57% on the above benchmarks, respectively. Code and pre-trained models for all the tasks are available at https://github.com/d-li14/involution.

Yikai Wang, Fuchun Sun, Duo Li et al.

We propose a general method to train a single convolutional neural network which is capable of switching image resolutions at inference. Thus the running speed can be selected to meet various computational resource limits. Networks trained with the proposed method are named Resolution Switchable Networks (RS-Nets). The basic training framework shares network parameters for handling images which differ in resolution, yet keeps separate batch normalization layers. Though it is parameter-efficient in design, it leads to inconsistent accuracy variations at different resolutions, for which we provide a detailed analysis from the aspect of the train-test recognition discrepancy. A multi-resolution ensemble distillation is further designed, where a teacher is learnt on the fly as a weighted ensemble over resolutions. Thanks to the ensemble and knowledge distillation, RS-Nets enjoy accuracy improvements at a wide range of resolutions compared with individually trained models. Extensive experiments on the ImageNet dataset are provided, and we additionally consider quantization problems. Code and models are available at https://github.com/yikaiw/RS-Nets.

Duo Li, Anbang Yao, Qifeng Chen

Despite their strong modeling capacities, Convolutional Neural Networks (CNNs) are often scale-sensitive. For enhancing the robustness of CNNs to scale variance, multi-scale feature fusion from different layers or filters attracts great attention among existing solutions, while the more granular kernel space is overlooked. We bridge this regret by exploiting multi-scale features in a finer granularity. The proposed convolution operation, named Poly-Scale Convolution (PSConv), mixes up a spectrum of dilation rates and tactfully allocate them in the individual convolutional kernels of each filter regarding a single convolutional layer. Specifically, dilation rates vary cyclically along the axes of input and output channels of the filters, aggregating features over a wide range of scales in a neat style. PSConv could be a drop-in replacement of the vanilla convolution in many prevailing CNN backbones, allowing better representation learning without introducing additional parameters and computational complexities. Comprehensive experiments on the ImageNet and MS COCO benchmarks validate the superior performance of PSConv. Code and models are available at https://github.com/d-li14/PSConv.

Duo Li, Anbang Yao, Qifeng Chen

Convolutional Neural Networks (CNNs) do not have a predictable recognition behavior with respect to the input resolution change. This prevents the feasibility of deployment on different input image resolutions for a specific model. To achieve efficient and flexible image classification at runtime, we employ meta learners to generate convolutional weights of main networks for various input scales and maintain privatized Batch Normalization layers per scale. For improved training performance, we further utilize knowledge distillation on the fly over model predictions based on different input resolutions. The learned meta network could dynamically parameterize main networks to act on input images of arbitrary size with consistently better accuracy compared to individually trained models. Extensive experiments on the ImageNet demonstrate that our method achieves an improved accuracy-efficiency trade-off during the adaptive inference process. By switching executable input resolutions, our method could satisfy the requirement of fast adaption in different resource-constrained environments. Code and models are available at https://github.com/d-li14/SAN.

Duo Li, Qifeng Chen

While the depth of modern Convolutional Neural Networks (CNNs) surpasses that of the pioneering networks with a significant margin, the traditional way of appending supervision only over the final classifier and progressively propagating gradient flow upstream remains the training mainstay. Seminal Deeply-Supervised Networks (DSN) were proposed to alleviate the difficulty of optimization arising from gradient flow through a long chain. However, it is still vulnerable to issues including interference to the hierarchical representation generation process and inconsistent optimization objectives, as illustrated theoretically and empirically in this paper. Complementary to previous training strategies, we propose Dynamic Hierarchical Mimicking, a generic feature learning mechanism, to advance CNN training with enhanced generalization ability. Partially inspired by DSN, we fork delicately designed side branches from the intermediate layers of a given neural network. Each branch can emerge from certain locations of the main branch dynamically, which not only retains representation rooted in the backbone network but also generates more diverse representations along its own pathway. We go one step further to promote multi-level interactions among different branches through an optimization formula with probabilistic prediction matching losses, thus guaranteeing a more robust optimization process and better representation ability. Experiments on both category and instance recognition tasks demonstrate the substantial improvements of our proposed method over its corresponding counterparts using diverse state-of-the-art CNN architectures. Code and models are publicly available at https://github.com/d-li14/DHM

Duo Li, Aojun Zhou, Anbang Yao

MobileNets, a class of top-performing convolutional neural network architectures in terms of accuracy and efficiency trade-off, are increasingly used in many resourceaware vision applications. In this paper, we present Harmonious Bottleneck on two Orthogonal dimensions (HBO), a novel architecture unit, specially tailored to boost the accuracy of extremely lightweight MobileNets at the level of less than 40 MFLOPs. Unlike existing bottleneck designs that mainly focus on exploring the interdependencies among the channels of either groupwise or depthwise convolutional features, our HBO improves bottleneck representation while maintaining similar complexity via jointly encoding the feature interdependencies across both spatial and channel dimensions. It has two reciprocal components, namely spatial contraction-expansion and channel expansion-contraction, nested in a bilaterally symmetric structure. The combination of two interdependent transformations performing on orthogonal dimensions of feature maps enhances the representation and generalization ability of our proposed module, guaranteeing compelling performance with limited computational resource and power. By replacing the original bottlenecks in MobileNetV2 backbone with HBO modules, we construct HBONets which are evaluated on ImageNet classification, PASCAL VOC object detection and Market-1501 person re-identification. Extensive experiments show that with the severe constraint of computational budget our models outperform MobileNetV2 counterparts by remarkable margins of at most 6.6%, 6.3% and 5.0% on the above benchmarks respectively. Code and pretrained models are available at https://github.com/d-li14/HBONet.

Duo Li, Zuhao Yang, Xiaoqin Zhang et al.

Visual token pruning aims to compress and prune redundant visual tokens which play a critical role in efficient inference with large vision-language models (LVLMs). However, most existing work estimates visual redundancy using a single metric, such as cross-modal attention or visual token similarity. We show that visual token diversity and task-specific token relevance are two crucial yet orthogonal factors that complement each other in conveying useful information and should therefore be treated separately for more effective visual token pruning. Building upon this insight, we design TODRE, a two-stage and training-free framework that incorporates Token Diversity and task RElevance for effective token compression and efficient LVLM inference. Instead of pruning redundant tokens, we introduce a greedy max-sum diversification algorithm that selects and retains a subset of diverse and representative visual tokens after the vision encoder. On top of that, ToDRE leverages an "information migration" mechanism to eliminate task-irrelevant visual tokens within certain decoder layers of large language model(LLM) to further improve token pruning and LVLM inference. Extensive experiments show that ToDRE prunes 90% of visual tokens after the vision encoder as well as all visual tokens in certain LLM decoder layers, leading to a 2.6x speed-up in total inference time while maintaining 95.0% model performance plus excellent model compatibility.

Duo Li, Zuhao Yang, Xiaoqin Zhang et al.

Discrete diffusion-based multimodal large language models (dMLLMs) have emerged as a promising alternative to autoregressive MLLMs thanks to their advantages in parallel decoding and bidirectional context modeling, but most existing dMLLMs incur significant computational overhead during inference due to the full-sequence attention computation in each denoising step. Pioneer studies attempt to resolve this issue from a modality-agnostic perspective via key-value cache optimization or efficient sampling but most of them overlook modality-specific visual token redundancy. In this work, we conduct a comprehensive study on how visual token redundancy evolves with different dMLLM architectures and tasks and how visual token pruning affects dMLLM responses and efficiency. Specifically, our study reveals that visual redundancy emerges only in from-scratch dMLLMs while handling long-answer tasks. In addition, we validate that visual token pruning introduces non-negligible information loss in dMLLMs and only from-scratch dMLLMs can recover the lost information progressively during late denoising steps. Furthermore, our study shows that layer-skipping is promising for accelerating AR-to-diffusion dMLLMs, whereas progressive or late-step pruning is more effective for from-scratch dMLLMs. Overall, this work offers a new perspective on efficiency optimization for dMLLMs, greatly advancing their applicability across various multimodal understanding tasks.

Wenhao Jiang, Duo Li, Menghan Hu et al.

In the field of autonomous driving, end-to-end deep learning models show great potential by learning driving decisions directly from sensor data. However, training these models requires large amounts of labeled data, which is time-consuming and expensive. Considering that the real-world driving data exhibits a long-tailed distribution where simple scenarios constitute a majority part of the data, we are thus inspired to identify the most challenging scenarios within it. Subsequently, we can efficiently improve the performance of the model by training with the selected data of the highest value. Prior research has focused on the selection of valuable data by empirically designed strategies. However, manually designed methods suffer from being less generalizable to new data distributions. Observing that the BEV (Bird's Eye View) features in end-to-end models contain all the information required to represent the scenario, we propose an active learning framework that relies on these vectorized scene-level features, called SEAD. The framework selects initial data based on driving-environmental information and incremental data based on BEV features. Experiments show that we only need 30\% of the nuScenes training data to achieve performance close to what can be achieved with the full dataset. The source code will be released.

Duo Li, Guimei Cao, Yunlu Xu et al.

In the SSLAD-Track 3B challenge on continual learning, we propose the method of COntinual Learning with Transformer (COLT). We find that transformers suffer less from catastrophic forgetting compared to convolutional neural network. The major principle of our method is to equip the transformer based feature extractor with old knowledge distillation and head expanding strategies to compete catastrophic forgetting. In this report, we first introduce the overall framework of continual learning for object detection. Then, we analyse the key elements' effect on withstanding catastrophic forgetting in our solution. Our method achieves 70.78 mAP on the SSLAD-Track 3B challenge test set.

Duo Li, Shang-Hua Gao

In recent years, the connections between deep residual networks and first-order Ordinary Differential Equations (ODEs) have been disclosed. In this work, we further bridge the deep neural architecture design with the second-order ODEs and propose a novel reversible neural network, termed as m-RevNet, that is characterized by inserting momentum update to residual blocks. The reversible property allows us to perform backward pass without access to activation values of the forward pass, greatly relieving the storage burden during training. Furthermore, the theoretical foundation based on second-order ODEs grants m-RevNet with stronger representational power than vanilla residual networks, which potentially explains its performance gains. For certain learning scenarios, we analytically and empirically reveal that our m-RevNet succeeds while standard ResNet fails. Comprehensive experiments on various image classification and semantic segmentation benchmarks demonstrate the superiority of our m-RevNet over ResNet, concerning both memory efficiency and recognition performance.

Lei Zhu, Qi She, Duo Li et al.

The nonlocal-based blocks are designed for capturing long-range spatial-temporal dependencies in computer vision tasks. Although having shown excellent performance, they still lack the mechanism to encode the rich, structured information among elements in an image or video. In this paper, to theoretically analyze the property of these nonlocal-based blocks, we provide a new perspective to interpret them, where we view them as a set of graph filters generated on a fully-connected graph. Specifically, when choosing the Chebyshev graph filter, a unified formulation can be derived for explaining and analyzing the existing nonlocal-based blocks (e.g., nonlocal block, nonlocal stage, double attention block). Furthermore, by concerning the property of spectral, we propose an efficient and robust spectral nonlocal block, which can be more robust and flexible to catch long-range dependencies when inserted into deep neural networks than the existing nonlocal blocks. Experimental results demonstrate the clear-cut improvements and practical applicabilities of our method on image classification, action recognition, semantic segmentation, and person re-identification tasks.

Dengsheng Chen, Haowen Deng, Jun Li et al.

Recently, various convolutions based on continuous or discrete kernels for point cloud processing have been widely studied, and achieve impressive performance in many applications, such as shape classification, scene segmentation and so on. However, they still suffer from some drawbacks. For continuous kernels, the inaccurate estimation of the kernel weights constitutes a bottleneck for further improving the performance; while for discrete ones, the kernels represented as the points located in the 3D space are lack of rich geometry information. In this work, rather than defining a continuous or discrete kernel, we directly embed convolutional kernels into the learnable potential fields, giving rise to potential convolution. It is convenient for us to define various potential functions for potential convolution which can generalize well to a wide range of tasks. Specifically, we provide two simple yet effective potential functions via point-wise convolution operations. Comprehensive experiments demonstrate the effectiveness of our method, which achieves superior performance on the popular 3D shape classification and scene segmentation benchmarks compared with other state-of-the-art point convolution methods.

Lei Zhu, Qi She, Bin Zhang et al.

Superpixel is generated by automatically clustering pixels in an image into hundreds of compact partitions, which is widely used to perceive the object contours for its excellent contour adherence. Although some works use the Convolution Neural Network (CNN) to generate high-quality superpixel, we challenge the design principles of these networks, specifically for their dependence on manual labels and excess computation resources, which limits their flexibility compared with the traditional unsupervised segmentation methods. We target at redefining the CNN-based superpixel segmentation as a lifelong clustering task and propose an unsupervised CNN-based method called LNS-Net. The LNS-Net can learn superpixel in a non-iterative and lifelong manner without any manual labels. Specifically, a lightweight feature embedder is proposed for LNS-Net to efficiently generate the cluster-friendly features. With those features, seed nodes can be automatically assigned to cluster pixels in a non-iterative way. Additionally, our LNS-Net can adapt the sequentially lifelong learning by rescaling the gradient of weight based on both channel and spatial context to avoid overfitting. Experiments show that the proposed LNS-Net achieves significantly better performance on three benchmarks with nearly ten times lower complexity compared with other state-of-the-art methods.

Duo Li, Qifeng Chen

In this paper, we build upon the weakly-supervised generation mechanism of intermediate attention maps in any convolutional neural networks and disclose the effectiveness of attention modules more straightforwardly to fully exploit their potential. Given an existing neural network equipped with arbitrary attention modules, we introduce a meta critic network to evaluate the quality of attention maps in the main network. Due to the discreteness of our designed reward, the proposed learning method is arranged in a reinforcement learning setting, where the attention actors and recurrent critics are alternately optimized to provide instant critique and revision for the temporary attention representation, hence coined as Deep REinforced Attention Learning (DREAL). It could be applied universally to network architectures with different types of attention modules and promotes their expressive ability by maximizing the relative gain of the final recognition performance arising from each individual attention module, as demonstrated by extensive experiments on both category and instance recognition benchmarks.