Yibo Yang

Jianzong Wu, Xiangtai Li, Shilin Xu et al.

In the field of visual scene understanding, deep neural networks have made impressive advancements in various core tasks like segmentation, tracking, and detection. However, most approaches operate on the close-set assumption, meaning that the model can only identify pre-defined categories that are present in the training set. Recently, open vocabulary settings were proposed due to the rapid progress of vision language pre-training. These new approaches seek to locate and recognize categories beyond the annotated label space. The open vocabulary approach is more general, practical, and effective compared to weakly supervised and zero-shot settings. This paper provides a thorough review of open vocabulary learning, summarizing and analyzing recent developments in the field. In particular, we begin by comparing it to related concepts such as zero-shot learning, open-set recognition, and out-of-distribution detection. Then, we review several closely related tasks in the case of segmentation and detection, including long-tail problems, few-shot, and zero-shot settings. For the method survey, we first present the basic knowledge of detection and segmentation in close-set as the preliminary knowledge. Next, we examine various scenarios in which open vocabulary learning is used, identifying common design elements and core ideas. Then, we compare the recent detection and segmentation approaches in commonly used datasets and benchmarks. Finally, we conclude with insights, issues, and discussions regarding future research directions. To our knowledge, this is the first comprehensive literature review of open vocabulary learning. We keep tracing related works at https://github.com/jianzongwu/Awesome-Open-Vocabulary.

Zhisheng Zhong, Jiequan Cui, Yibo Yang et al.

A recent study has shown a phenomenon called neural collapse in that the within-class means of features and the classifier weight vectors converge to the vertices of a simplex equiangular tight frame at the terminal phase of training for classification. In this paper, we explore the corresponding structures of the last-layer feature centers and classifiers in semantic segmentation. Based on our empirical and theoretical analysis, we point out that semantic segmentation naturally brings contextual correlation and imbalanced distribution among classes, which breaks the equiangular and maximally separated structure of neural collapse for both feature centers and classifiers. However, such a symmetric structure is beneficial to discrimination for the minor classes. To preserve these advantages, we introduce a regularizer on feature centers to encourage the network to learn features closer to the appealing structure in imbalanced semantic segmentation. Experimental results show that our method can bring significant improvements on both 2D and 3D semantic segmentation benchmarks. Moreover, our method ranks 1st and sets a new record (+6.8% mIoU) on the ScanNet200 test leaderboard. Code will be available at https://github.com/dvlab-research/Imbalanced-Learning.

Xiangtai Li, Shilin Xu, Yibo Yang et al.

Panoptic Part Segmentation (PPS) unifies panoptic and part segmentation into one task. Previous works utilize separate approaches to handle things, stuff, and part predictions without shared computation and task association. We aim to unify these tasks at the architectural level, designing the first end-to-end unified framework, Panoptic-PartFormer. Moreover, we find the previous metric PartPQ biases to PQ. To handle both issues, we first design a meta-architecture that decouples part features and things/stuff features, respectively. We model things, stuff, and parts as object queries and directly learn to optimize all three forms of prediction as a unified mask prediction and classification problem. We term our model as Panoptic-PartFormer. Second, we propose a new metric Part-Whole Quality (PWQ), better to measure this task from pixel-region and part-whole perspectives. It also decouples the errors for part segmentation and panoptic segmentation. Third, inspired by Mask2Former, based on our meta-architecture, we propose Panoptic-PartFormer++ and design a new part-whole cross-attention scheme to boost part segmentation qualities further. We design a new part-whole interaction method using masked cross attention. Finally, extensive ablation studies and analysis demonstrate the effectiveness of both Panoptic-PartFormer and Panoptic-PartFormer++. Compared with previous Panoptic-PartFormer, our Panoptic-PartFormer++ achieves 2% PartPQ and 3% PWQ improvements on the Cityscapes PPS dataset and 5% PartPQ on the Pascal Context PPS dataset. On both datasets, Panoptic-PartFormer++ achieves new state-of-the-art results. Our models can serve as a strong baseline and aid future research in PPS. The source code and trained models will be available at~\url{https://github.com/lxtGH/Panoptic-PartFormer}.

Xiangtai Li, Jiangning Zhang, Yibo Yang et al.

In this paper, we focus on exploring effective methods for faster and accurate semantic segmentation. A common practice to improve the performance is to attain high-resolution feature maps with strong semantic representation. Two strategies are widely used: atrous convolutions and feature pyramid fusion, while both are either computationally intensive or ineffective. Inspired by the Optical Flow for motion alignment between adjacent video frames, we propose a Flow Alignment Module (FAM) to learn \textit{Semantic Flow} between feature maps of adjacent levels and broadcast high-level features to high-resolution features effectively and efficiently. Furthermore, integrating our FAM to a standard feature pyramid structure exhibits superior performance over other real-time methods, even on lightweight backbone networks, such as ResNet-18 and DFNet. Then to further speed up the inference procedure, we also present a novel Gated Dual Flow Alignment Module to directly align high-resolution feature maps and low-resolution feature maps where we term the improved version network as SFNet-Lite. Extensive experiments are conducted on several challenging datasets, where results show the effectiveness of both SFNet and SFNet-Lite. In particular, when using Cityscapes test set, the SFNet-Lite series achieve 80.1 mIoU while running at 60 FPS using ResNet-18 backbone and 78.8 mIoU while running at 120 FPS using STDC backbone on RTX-3090. Moreover, we unify four challenging driving datasets into one large dataset, which we named Unified Driving Segmentation (UDS) dataset. It contains diverse domain and style information. We benchmark several representative works on UDS. Both SFNet and SFNet-Lite still achieve the best speed and accuracy trade-off on UDS, which serves as a strong baseline in such a challenging setting. The code and models are publicly available at https://github.com/lxtGH/SFSegNets.

Zekun Li, Zhengyang Geng, Zhao Kang et al.

Reference-based line-art colorization is a challenging task in computer vision. The color, texture, and shading are rendered based on an abstract sketch, which heavily relies on the precise long-range dependency modeling between the sketch and reference. Popular techniques to bridge the cross-modal information and model the long-range dependency employ the attention mechanism. However, in the context of reference-based line-art colorization, several techniques would intensify the existing training difficulty of attention, for instance, self-supervised training protocol and GAN-based losses. To understand the instability in training, we detect the gradient flow of attention and observe gradient conflict among attention branches. This phenomenon motivates us to alleviate the gradient issue by preserving the dominant gradient branch while removing the conflict ones. We propose a novel attention mechanism using this training strategy, Stop-Gradient Attention (SGA), outperforming the attention baseline by a large margin with better training stability. Compared with state-of-the-art modules in line-art colorization, our approach demonstrates significant improvements in Fréchet Inception Distance (FID, up to 27.21%) and structural similarity index measure (SSIM, up to 25.67%) on several benchmarks. The code of SGA is available at https://github.com/kunkun0w0/SGA .

Yibo Yang, Haobo Yuan, Xiangtai Li et al.

How to enable learnability for new classes while keeping the capability well on old classes has been a crucial challenge for class incremental learning. Beyond the normal case, long-tail class incremental learning and few-shot class incremental learning are also proposed to consider the data imbalance and data scarcity, respectively, which are common in real-world implementations and further exacerbate the well-known problem of catastrophic forgetting. Existing methods are specifically proposed for one of the three tasks. In this paper, we offer a unified solution to the misalignment dilemma in the three tasks. Concretely, we propose neural collapse terminus that is a fixed structure with the maximal equiangular inter-class separation for the whole label space. It serves as a consistent target throughout the incremental training to avoid dividing the feature space incrementally. For CIL and LTCIL, we further propose a prototype evolving scheme to drive the backbone features into our neural collapse terminus smoothly. Our method also works for FSCIL with only minor adaptations. Theoretical analysis indicates that our method holds the neural collapse optimality in an incremental fashion regardless of data imbalance or data scarcity. We also design a generalized case where we do not know the total number of classes and whether the data distribution is normal, long-tail, or few-shot for each coming session, to test the generalizability of our method. Extensive experiments with multiple datasets are conducted to demonstrate the effectiveness of our unified solution to all the three tasks and the generalized case.

Yibo Yang, Hong Wang, Haobo Yuan et al.

Automated machine learning has been widely explored to reduce human efforts in designing neural architectures and looking for proper hyperparameters. In the domain of neural initialization, however, similar automated techniques have rarely been studied. Most existing initialization methods are handcrafted and highly dependent on specific architectures. In this paper, we propose a differentiable quantity, named GradCosine, with theoretical insights to evaluate the initial state of a neural network. Specifically, GradCosine is the cosine similarity of sample-wise gradients with respect to the initialized parameters. By analyzing the sample-wise optimization landscape, we show that both the training and test performance of a network can be improved by maximizing GradCosine under gradient norm constraint. Based on this observation, we further propose the neural initialization optimization (NIO) algorithm. Generalized from the sample-wise analysis into the real batch setting, NIO is able to automatically look for a better initialization with negligible cost compared with the training time. With NIO, we improve the classification performance of a variety of neural architectures on CIFAR-10, CIFAR-100, and ImageNet. Moreover, we find that our method can even help to train large vision Transformer architecture without warmup.

Yuxuan Du, Yibo Yang, Tongliang Liu et al.

Understanding the dynamics of large quantum systems is hindered by the curse of dimensionality. Statistical learning offers new possibilities in this regime by neural-network protocols and classical shadows, while both methods have limitations: the former is plagued by the predictive uncertainty and the latter lacks the generalization ability. Here we propose a data-centric learning paradigm combining the strength of these two approaches to facilitate diverse quantum system learning (QSL) tasks. Particularly, our paradigm utilizes classical shadows along with other easily obtainable information of quantum systems to create the training dataset, which is then learnt by neural networks to unveil the underlying mapping rule of the explored QSL problem. Capitalizing on the generalization power of neural networks, this paradigm can be trained offline and excel at predicting previously unseen systems at the inference stage, even with few state copies. Besides, it inherits the characteristic of classical shadows, enabling memory-efficient storage and faithful prediction. These features underscore the immense potential of the proposed data-centric approach in discovering novel and large-scale quantum systems. For concreteness, we present the instantiation of our paradigm in quantum state tomography and direct fidelity estimation tasks and conduct numerical analysis up to 60 qubits. Our work showcases the profound prospects of data-centric artificial intelligence to advance QSL in a faithful and generalizable manner.

Xiaojie Li, Yibo Yang, Jianlong Wu et al.

Few-shot class-incremental learning (FSCIL) aims to incrementally learn novel classes from limited examples while preserving knowledge of previously learned classes. Existing methods face a critical dilemma: static architectures rely on a fixed parameter space to learn from data that arrive sequentially, prone to overfitting to the current session, while dynamic architectures require the expansion of the parameter space continually, leading to increased complexity. In this study, we explore the potential of Selective State Space Models (SSMs) for FSCIL. Mamba leverages its input-dependent parameters to dynamically adjust its processing patterns and generate content-aware scan patterns within a fixed architecture. This enables it to configure distinct processing for base and novel classes, effectively preserving existing knowledge while adapting to new ones. To leverage Mamba's potential for FSCIL, we design two key modules: First, we propose a dual selective SSM projector that dynamically adjusts the projection parameters based on the intermediate features for dynamic adaptation. The dual-design structurally decouples base and novel class processing with a frozen base branch, employing a frozen base branch to maintain robust base-class features and a dynamic incremental branch that adaptively learns distinctive feature shifts for novel classes. Second, we develop a class-sensitive selective scan mechanism to guide dynamic adaptation of the incremental branch. It minimizes the disruption to base-class representations caused by training on novel data, and meanwhile, forces the selective scan to perform in distinct patterns between base and novel classes. Extensive experiments on miniImageNet, CUB-200, and CIFAR-100 demonstrate that Mamba-FSCIL achieves state-of-the-art performance. The code is available at https://github.com/xiaojieli0903/Mamba-FSCIL.

Liang Xie, Yibo Yang, Wenxiao Wang et al.

Compared to 2D images, 3D point clouds are much more sensitive to rotations. We expect the point features describing certain patterns to keep invariant to the rotation transformation. There are many recent SOTA works dedicated to rotation-invariant learning for 3D point clouds. However, current rotation-invariant methods lack generalizability on the point clouds in the open scenes due to the reliance on the global distribution, \ie the global scene and backgrounds. Considering that the output activation is a function of the pattern and its orientation, we need to eliminate the effect of the orientation.In this paper, inspired by the idea that the network weights can be considered a set of points distributed in the same 3D space as the input points, we propose Weight-Feature Alignment (WFA) to construct a local Invariant Reference Frame (IRF) via aligning the features with the principal axes of the network weights. Our WFA algorithm provides a general solution for the point clouds of all scenes. WFA ensures the model achieves the target that the response activity is a necessary and sufficient condition of the pattern matching degree. Practically, we perform experiments on the point clouds of both single objects and open large-range scenes. The results suggest that our method almost bridges the gap between rotation invariance learning and normal methods.

Yibo Yang, Stephan Eckstein, Marcel Nutz et al.

In the theory of lossy compression, the rate-distortion (R-D) function $R(D)$ describes how much a data source can be compressed (in bit-rate) at any given level of fidelity (distortion). Obtaining $R(D)$ for a given data source establishes the fundamental performance limit for all compression algorithms. We propose a new method to estimate $R(D)$ from the perspective of optimal transport. Unlike the classic Blahut--Arimoto algorithm which fixes the support of the reproduction distribution in advance, our Wasserstein gradient descent algorithm learns the support of the optimal reproduction distribution by moving particles. We prove its local convergence and analyze the sample complexity of our R-D estimator based on a connection to entropic optimal transport. Experimentally, we obtain comparable or tighter bounds than state-of-the-art neural network methods on low-rate sources while requiring considerably less tuning and computation effort. We also highlight a connection to maximum-likelihood deconvolution and introduce a new class of sources that can be used as test cases with known solutions to the R-D problem.

Robin Wang, Yibo Yang, Dacheng Tao

Point cloud classifiers with rotation robustness have been widely discussed in the 3D deep learning community. Most proposed methods either use rotation invariant descriptors as inputs or try to design rotation equivariant networks. However, robust models generated by these methods have limited performance under clean aligned datasets due to modifications on the original classifiers or input space. In this study, for the first time, we show that the rotation robustness of point cloud classifiers can also be acquired via adversarial training with better performance on both rotated and clean datasets. Specifically, our proposed framework named ART-Point regards the rotation of the point cloud as an attack and improves rotation robustness by training the classifier on inputs with Adversarial RoTations. We contribute an axis-wise rotation attack that uses back-propagated gradients of the pre-trained model to effectively find the adversarial rotations. To avoid model over-fitting on adversarial inputs, we construct rotation pools that leverage the transferability of adversarial rotations among samples to increase the diversity of training data. Moreover, we propose a fast one-step optimization to efficiently reach the final robust model. Experiments show that our proposed rotation attack achieves a high success rate and ART-Point can be used on most existing classifiers to improve the rotation robustness while showing better performance on clean datasets than state-of-the-art methods.

Peifeng Gao, Qianqian Xu, Yibo Yang et al.

Neural Collapse (NC) is a well-known phenomenon of deep neural networks in the terminal phase of training (TPT). It is characterized by the collapse of features and classifier into a symmetrical structure, known as simplex equiangular tight frame (ETF). While there have been extensive studies on optimization characteristics showing the global optimality of neural collapse, little research has been done on the generalization behaviors during the occurrence of NC. Particularly, the important phenomenon of generalization improvement during TPT has been remaining in an empirical observation and lacking rigorous theoretical explanation. In this paper, we establish the connection between the minimization of CE and a multi-class SVM during TPT, and then derive a multi-class margin generalization bound, which provides a theoretical explanation for why continuing training can still lead to accuracy improvement on test set, even after the train accuracy has reached 100%. Additionally, our further theoretical results indicate that different alignment between labels and features in a simplex ETF can result in varying degrees of generalization improvement, despite all models reaching NC and demonstrating similar optimization performance on train set. We refer to this newly discovered property as "non-conservative generalization". In experiments, we also provide empirical observations to verify the indications suggested by our theoretical results.

Xiaojie Li, Yibo Yang, Xiangtai Li et al.

Self-supervised learning has achieved remarkable success in acquiring high-quality representations from unlabeled data. The widely adopted contrastive learning framework aims to learn invariant representations by minimizing the distance between positive views originating from the same image. However, existing techniques to construct positive views highly rely on manual transformations, resulting in limited diversity and potentially false positive pairs. To tackle these challenges, we present GenView, a controllable framework that augments the diversity of positive views leveraging the power of pretrained generative models while preserving semantics. We develop an adaptive view generation method that dynamically adjusts the noise level in sampling to ensure the preservation of essential semantic meaning while introducing variability. Additionally, we introduce a quality-driven contrastive loss, which assesses the quality of positive pairs by considering both foreground similarity and background diversity. This loss prioritizes the high-quality positive pairs we construct while reducing the influence of low-quality pairs, thereby mitigating potential semantic inconsistencies introduced by generative models and aggressive data augmentation. Thanks to the improved positive view quality and the quality-driven contrastive loss, GenView significantly improves self-supervised learning across various tasks. For instance, GenView improves MoCov2 performance by 2.5%/2.2% on ImageNet linear/semi-supervised classification. Moreover, GenView even performs much better than naively augmenting the ImageNet dataset with Laion400M or ImageNet21K. Code: https://github.com/xiaojieli0903/genview.

Boyan Gao, Bo Zhao, Shreyank N Gowda et al.

Dataset condensation aims to synthesize datasets with a few representative samples that can effectively represent the original datasets. This enables efficient training and produces models with performance close to those trained on the original sets. Most existing dataset condensation methods conduct dataset learning under the bilevel (inner- and outer-loop) based optimization. However, the preceding methods perform with limited dataset generalization due to the notoriously complicated loss landscape and expensive time-space complexity of the inner-loop unrolling of bilevel optimization. These issues deteriorate when the datasets are learned via matching the trajectories of networks trained on the real and synthetic datasets with a long horizon inner-loop. To address these issues, we introduce Sharpness-Aware Trajectory Matching (SATM), which enhances the generalization capability of learned synthetic datasets by optimising the sharpness of the loss landscape and objective simultaneously. Moreover, our approach is coupled with an efficient hypergradient approximation that is mathematically well-supported and straightforward to implement along with controllable computational overhead. Empirical evaluations of SATM demonstrate its effectiveness across various applications, including in-domain benchmarks and out-of-domain settings. Moreover, its easy-to-implement properties afford flexibility, allowing it to integrate with other advanced sharpness-aware minimizers. Our code will be released.

Shilin Xu, Haobo Yuan, Qingyu Shi et al.

Recent segmentation methods, which adopt large-scale data training and transformer architecture, aim to create one foundation model that can perform multiple tasks. However, most of these methods rely on heavy encoder and decoder frameworks, hindering their performance in real-time scenarios. To explore real-time segmentation, recent advancements primarily focus on semantic segmentation within specific environments, such as autonomous driving. However, they often overlook the generalization ability of these models across diverse scenarios. Therefore, to fill this gap, this work explores a novel real-time segmentation setting called real-time multi-purpose segmentation. It contains three fundamental sub-tasks: interactive segmentation, panoptic segmentation, and video instance segmentation. Unlike previous methods, which use a specific design for each task, we aim to use only a single end-to-end model to accomplish all these tasks in real-time. To meet real-time requirements and balance multi-task learning, we present a novel dynamic convolution-based method, Real-Time Multi-Purpose SAM (RMP-SAM). It contains an efficient encoder and an efficient decoupled adapter to perform prompt-driven decoding. Moreover, we further explore different training strategies and one new adapter design to boost co-training performance further. We benchmark several strong baselines by extending existing works to support our multi-purpose segmentation. Extensive experiments demonstrate that RMP-SAM is effective and generalizes well on proposed benchmarks and other specific semantic tasks. Our implementation of RMP-SAM achieves the optimal balance between accuracy and speed for these tasks.Our code and model are available at https://github.com/xushilin1/RAP-SAM/.

Yibo Yang, Stephan Mandt

Rate-distortion (R-D) function, a key quantity in information theory, characterizes the fundamental limit of how much a data source can be compressed subject to a fidelity criterion, by any compression algorithm. As researchers push for ever-improving compression performance, establishing the R-D function of a given data source is not only of scientific interest, but also sheds light on the possible room for improving compression algorithms. Previous work on this problem relied on distributional assumptions on the data source (Gibson, 2017) or only applied to discrete data (Blahut, 1972; Arimoto, 1972). By contrast, this paper makes the first attempt at an algorithm for sandwiching the R-D function of a general (not necessarily discrete) source requiring only i.i.d. data samples. We estimate R-D sandwich bounds for a variety of artificial and real-world data sources, in settings far beyond the feasibility of any known method, and shed light on the optimality of neural data compression (Ballé et al., 2021; Yang et al., 2022). Our R-D upper bound on natural images indicates theoretical room for improving state-of-the-art image compression methods by at least one dB in PSNR at various bitrates. Our data and code can be found at https://github.com/mandt-lab/empirical-RD-sandwich.

Yibo Yang, Sihao Liu, Chuan Rao et al.

Conventional low-rank adaptation methods build adapters without considering data context, leading to sub-optimal fine-tuning performance and severe forgetting of inherent world knowledge. In this paper, we propose context-oriented decomposition adaptation (CorDA), a novel method that initializes adapters in a task-aware manner. Concretely, we develop context-oriented singular value decomposition, where we collect covariance matrices of input activations for each linear layer using sampled data from the target task, and apply SVD to the product of weight matrix and its corresponding covariance matrix. By doing so, the task-specific capability is compacted into the principal components. Thanks to the task awareness, our method enables two optional adaptation modes, knowledge-preserved mode (KPM) and instruction-previewed mode (IPM), providing flexibility to choose between freezing the principal components to preserve their associated knowledge or adapting them to better learn a new task. We further develop CorDA++ by deriving a metric that reflects the compactness of task-specific principal components, and then introducing dynamic covariance selection and dynamic rank allocation strategies based on the same metric. The two strategies provide each layer with the most representative covariance matrix and a proper rank allocation. Experimental results show that CorDA++ outperforms CorDA by a significant margin. CorDA++ in KPM not only achieves better fine-tuning performance than LoRA, but also mitigates the forgetting of pre-trained knowledge in both large language models and vision language models. For IPM, our method exhibits faster convergence, \emph{e.g.,} 4.5x speedup over QLoRA, and improves adaptation performance in various scenarios, outperforming strong baseline methods. Our method has been integrated into the PEFT library developed by Hugging Face.

Yibo Yang, Justus C. Will, Stephan Mandt

Diffusion probabilistic models have achieved mainstream success in many generative modeling tasks, from image generation to inverse problem solving. A distinct feature of these models is that they correspond to deep hierarchical latent variable models optimizing a variational evidence lower bound (ELBO) on the data likelihood. Drawing on a basic connection between likelihood modeling and compression, we explore the potential of diffusion models for progressive coding, resulting in a sequence of bits that can be incrementally transmitted and decoded with progressively improving reconstruction quality. Unlike prior work based on Gaussian diffusion or conditional diffusion models, we propose a new form of diffusion model with uniform noise in the forward process, whose negative ELBO corresponds to the end-to-end compression cost using universal quantization. We obtain promising first results on image compression, achieving competitive rate-distortion and rate-realism results on a wide range of bit-rates with a single model, bringing neural codecs a step closer to practical deployment.

Sihao Liu, Yibo Yang, Xiaojie Li et al.

Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data. To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics. We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results.

Peiqi Wang, ShengYun Peng, Xuewen Zhang et al.

This work investigates the optimal allocation of inference compute across three key scaling factors in video vision language models: language model size, frame count, and the number of visual tokens per frame. While prior works typically focuses on optimizing model efficiency or improving performance without considering resource constraints, we instead identify optimal model configuration under fixed inference compute budgets. We conduct large-scale training sweeps and careful parametric modeling of task performance to identify the inference compute-optimal frontier. Our experiments reveal how task performance depends on scaling factors and finetuning data size, as well as how changes in data size shift the compute-optimal frontier. These findings translate to practical tips for selecting these scaling factors.

Zhongzhu Zhou, Yibo Yang, Ziyan Chen et al.

Policy gradient (PG) methods in reinforcement learning frequently utilize deep neural networks (DNNs) to learn a shared backbone of feature representations used to compute likelihoods in an action selection layer. Numerous studies have been conducted on the convergence and global optima of policy networks, but few have analyzed representational structures of those underlying networks. While training an optimal policy DNN, we observed that under certain constraints, a gentle structure resembling neural collapse, which we refer to as Action Collapse (AC), emerges. This suggests that 1) the state-action activations (i.e. last-layer features) sharing the same optimal actions collapse towards those optimal actions respective mean activations; 2) the variability of activations sharing the same optimal actions converges to zero; 3) the weights of action selection layer and the mean activations collapse to a simplex equiangular tight frame (ETF). Our early work showed those aforementioned constraints to be necessary for these observations. Since the collapsed ETF of optimal policy DNNs maximally separates the pair-wise angles of all actions in the state-action space, we naturally raise a question: can we learn an optimal policy using an ETF structure as a (fixed) target configuration in the action selection layer? Our analytical proof shows that learning activations with a fixed ETF as action selection layer naturally leads to the AC. We thus propose the Action Collapse Policy Gradient (ACPG) method, which accordingly affixes a synthetic ETF as our action selection layer. ACPG induces the policy DNN to produce such an ideal configuration in the action selection layer while remaining optimal. Our experiments across various OpenAI Gym environments demonstrate that our technique can be integrated into any discrete PG methods and lead to favorable reward improvements more quickly and robustly.

Jiani Ni, He Zhao, Yibo Yang et al.

In recent years, deep neural networks (DNNs) have shown competitive results in many fields. Despite this success, they often suffer from poor calibration, especially in safety-critical scenarios such as autonomous driving and healthcare, where unreliable confidence estimates can lead to serious consequences. Recent studies have focused on improving calibration by modifying the classifier, yet such efforts remain limited. Moreover, most existing approaches overlook calibration errors caused by underconfidence, which can be equally detrimental. To address these challenges, we propose MaC-Cal, a novel mask-based classifier calibration method that leverages stochastic sparsity to enhance the alignment between confidence and accuracy. MaC-Cal adopts a two-stage training scheme with adaptive sparsity, dynamically adjusting mask retention rates based on the deviation between confidence and accuracy. Extensive experiments show that MaC-Cal achieves superior calibration performance and robustness under data corruption, offering a practical and effective solution for reliable confidence estimation in DNNs.

Tuan Truong, Rithwik Sudharsan, Yibo Yang et al.

The site conditions that make astronomical observatories in space and on the ground so desirable -- cold and dark -- demand a physical remoteness that leads to limited data transmission capabilities. Such transmission limitations directly bottleneck the amount of data acquired and in an era of costly modern observatories, any improvements in lossless data compression has the potential scale to billions of dollars worth of additional science that can be accomplished on the same instrument. Traditional lossless methods for compressing astrophysical data are manually designed. Neural data compression, on the other hand, holds the promise of learning compression algorithms end-to-end from data and outperforming classical techniques by leveraging the unique spatial, temporal, and wavelength structures of astronomical images. This paper introduces AstroCompress: a neural compression challenge for astrophysics data, featuring four new datasets (and one legacy dataset) with 16-bit unsigned integer imaging data in various modes: space-based, ground-based, multi-wavelength, and time-series imaging. We provide code to easily access the data and benchmark seven lossless compression methods (three neural and four non-neural, including all practical state-of-the-art algorithms). Our results on lossless compression indicate that lossless neural compression techniques can enhance data collection at observatories, and provide guidance on the adoption of neural compression in scientific applications. Though the scope of this paper is restricted to lossless compression, we also comment on the potential exploration of lossy compression methods in future studies.

Boyan Gao, Xin Wang, Yibo Yang et al.

Large Language Models (LLMs) have demonstrated remarkable performance in real-world applications. However, adapting LLMs to novel tasks via fine-tuning often requires substantial training data and computational resources that are impractical in few-shot scenarios. Existing approaches, such as in-context learning and Parameter-Efficient Fine-Tuning (PEFT), face key limitations: in-context learning introduces additional inference computational overhead with limited performance gains, while PEFT models are prone to overfitting on the few demonstration examples. In this work, we reinterpret the forward pass of LLMs as an optimization process, a sequence of preconditioned gradient descent steps refining internal representations. Based on this connection, we propose Optimization-Inspired Few-Shot Adaptation (OFA), integrating a parameterization that learns preconditioners without introducing additional trainable parameters, and an objective that improves optimization efficiency by learning preconditioners based on a convergence bound, while simultaneously steering the optimization path toward the flat local minimum. Our method overcomes both issues of ICL-based and PEFT-based methods, and demonstrates superior performance over the existing methods on a variety of few-shot adaptation tasks in experiments.

Zitian Wang, Yue Liao, Kang Rong et al.

Preference alignment has emerged as an effective strategy to enhance the performance of Multimodal Large Language Models (MLLMs) following supervised fine-tuning. While existing preference alignment methods predominantly target hallucination factors, they overlook the factors essential for multi-modal comprehension capabilities, often narrowing their improvements on hallucination mitigation. To bridge this gap, we propose Instruction-oriented Preference Alignment (IPA), a scalable framework designed to automatically construct alignment preferences grounded in instruction fulfillment efficacy. Our method involves an automated preference construction coupled with a dedicated verification process that identifies instruction-oriented factors, avoiding significant variability in response representations. Additionally, IPA incorporates a progressive preference collection pipeline, further recalling challenging samples through model self-evolution and reference-guided refinement. Experiments conducted on Qwen2VL-7B demonstrate IPA's effectiveness across multiple benchmarks, including hallucination evaluation, visual question answering, and text understanding tasks, highlighting its capability to enhance general comprehension.

Yibo Yang, Xiaojie Li, Zhongzhu Zhou et al.

Current parameter-efficient fine-tuning (PEFT) methods build adapters widely agnostic of the context of downstream task to learn, or the context of important knowledge to maintain. As a result, there is often a performance gap compared to full-parameter fine-tuning, and meanwhile the fine-tuned model suffers from catastrophic forgetting of the pre-trained world knowledge. In this paper, we propose CorDA, a Context-oriented Decomposition Adaptation method that builds learnable task-aware adapters from weight decomposition oriented by the context of downstream task or the world knowledge to maintain. Concretely, we collect a few data samples, and perform singular value decomposition for each linear layer of a pre-trained LLM multiplied by the covariance matrix of the input activation using these samples. The inverse of the covariance matrix is multiplied with the decomposed components to reconstruct the original weights. By doing so, the context of the representative samples is captured through deciding the factorizing orientation. Our method enables two options, the knowledge-preserved adaptation and the instruction-previewed adaptation. For the former, we use question-answering samples to obtain the covariance matrices, and use the decomposed components with the smallest $r$ singular values to initialize a learnable adapter, with the others frozen such that the world knowledge is better preserved. For the latter, we use the instruction data from the fine-tuning task, such as math or coding, to orientate the decomposition and train the largest $r$ components that most correspond to the task to learn. We conduct extensive experiments on Math, Code, and Instruction Following tasks.

Yibo Yang, Xiaojie Li, Motasem Alfarra et al.

Relieving the reliance of neural network training on a global back-propagation (BP) has emerged as a notable research topic due to the biological implausibility and huge memory consumption caused by BP. Among the existing solutions, local learning optimizes gradient-isolated modules of a neural network with local errors and has been proved to be effective even on large-scale datasets. However, the reconciliation among local errors has never been investigated. In this paper, we first theoretically study non-greedy layer-wise training and show that the convergence cannot be assured when the local gradient in a module w.r.t. its input is not reconciled with the local gradient in the previous module w.r.t. its output. Inspired by the theoretical result, we further propose a local training strategy that successively regularizes the gradient reconciliation between neighboring modules without breaking gradient isolation or introducing any learnable parameters. Our method can be integrated into both local-BP and BP-free settings. In experiments, we achieve significant performance improvements compared to previous methods. Particularly, our method for CNN and Transformer architectures on ImageNet is able to attain a competitive performance with global BP, saving more than 40% memory consumption.

Fengxiang Bie, Yibo Yang, Zhongzhu Zhou et al.

Text-to-image generation (TTI) refers to the usage of models that could process text input and generate high fidelity images based on text descriptions. Text-to-image generation using neural networks could be traced back to the emergence of Generative Adversial Network (GAN), followed by the autoregressive Transformer. Diffusion models are one prominent type of generative model used for the generation of images through the systematic introduction of noises with repeating steps. As an effect of the impressive results of diffusion models on image synthesis, it has been cemented as the major image decoder used by text-to-image models and brought text-to-image generation to the forefront of machine-learning (ML) research. In the era of large models, scaling up model size and the integration with large language models have further improved the performance of TTI models, resulting the generation result nearly indistinguishable from real-world images, revolutionizing the way we retrieval images. Our explorative study has incentivised us to think that there are further ways of scaling text-to-image models with the combination of innovative model architectures and prediction enhancement techniques. We have divided the work of this survey into five main sections wherein we detail the frameworks of major literature in order to delve into the different types of text-to-image generation methods. Following this we provide a detailed comparison and critique of these methods and offer possible pathways of improvement for future work. In the future work, we argue that TTI development could yield impressive productivity improvements for creation, particularly in the context of the AIGC era, and could be extended to more complex tasks such as video generation and 3D generation.

Yibo Yang, Stephan Mandt, Lucas Theis

Neural compression is the application of neural networks and other machine learning methods to data compression. Recent advances in statistical machine learning have opened up new possibilities for data compression, allowing compression algorithms to be learned end-to-end from data using powerful generative models such as normalizing flows, variational autoencoders, diffusion probabilistic models, and generative adversarial networks. The present article aims to introduce this field of research to a broader machine learning audience by reviewing the necessary background in information theory (e.g., entropy coding, rate-distortion theory) and computer vision (e.g., image quality assessment, perceptual metrics), and providing a curated guide through the essential ideas and methods in the literature thus far.

Qianyu Zhou, Xiangtai Li, Lu He et al.

Detection Transformer (DETR) and Deformable DETR have been proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance as previous complex hand-crafted detectors. However, their performance on Video Object Detection (VOD) has not been well explored. In this paper, we present TransVOD, the first end-to-end video object detection system based on spatial-temporal Transformer architectures. The first goal of this paper is to streamline the pipeline of VOD, effectively removing the need for many hand-crafted components for feature aggregation, e.g., optical flow model, relation networks. Besides, benefited from the object query design in DETR, our method does not need complicated post-processing methods such as Seq-NMS. In particular, we present a temporal Transformer to aggregate both the spatial object queries and the feature memories of each frame. Our temporal transformer consists of two components: Temporal Query Encoder (TQE) to fuse object queries, and Temporal Deformable Transformer Decoder (TDTD) to obtain current frame detection results. These designs boost the strong baseline deformable DETR by a significant margin (3%-4% mAP) on the ImageNet VID dataset. Then, we present two improved versions of TransVOD including TransVOD++ and TransVOD Lite. The former fuses object-level information into object query via dynamic convolution while the latter models the entire video clips as the output to speed up the inference time. We give detailed analysis of all three models in the experiment part. In particular, our proposed TransVOD++ sets a new state-of-the-art record in terms of accuracy on ImageNet VID with 90.0% mAP. Our proposed TransVOD Lite also achieves the best speed and accuracy trade-off with 83.7% mAP while running at around 30 FPS on a single V100 GPU device.

Ruihan Yang, Yibo Yang, Joseph Marino et al.

While recent machine learning research has revealed connections between deep generative models such as VAEs and rate-distortion losses used in learned compression, most of this work has focused on images. In a similar spirit, we view recently proposed neural video coding algorithms through the lens of deep autoregressive and latent variable modeling. We present these codecs as instances of a generalized stochastic temporal autoregressive transform, and propose new avenues for further improvements inspired by normalizing flows and structured priors. We propose several architectures that yield state-of-the-art video compression performance on high-resolution video and discuss their tradeoffs and ablations. In particular, we propose (i) improved temporal autoregressive transforms, (ii) improved entropy models with structured and temporal dependencies, and (iii) variable bitrate versions of our algorithms. Since our improvements are compatible with a large class of existing models, we provide further evidence that the generative modeling viewpoint can advance the neural video coding field.

Yibo Yang, Antoine Blanchard, Themistoklis Sapsis et al.

We present a new type of acquisition functions for online decision making in multi-armed and contextual bandit problems with extreme payoffs. Specifically, we model the payoff function as a Gaussian process and formulate a novel type of upper confidence bound (UCB) acquisition function that guides exploration towards the bandits that are deemed most relevant according to the variability of the observed rewards. This is achieved by computing a tractable likelihood ratio that quantifies the importance of the output relative to the inputs and essentially acts as an \textit{attention mechanism} that promotes exploration of extreme rewards. We demonstrate the benefits of the proposed methodology across several synthetic benchmarks, as well as a realistic example involving noisy sensor network data. Finally, we provide a JAX library for efficient bandit optimization using Gaussian processes.

Yibo Yang, Shan You, Hongyang Li et al.

Most differentiable neural architecture search methods construct a super-net for search and derive a target-net as its sub-graph for evaluation. There exists a significant gap between the architectures in search and evaluation. As a result, current methods suffer from an inconsistent, inefficient, and inflexible search process. In this paper, we introduce EnTranNAS that is composed of Engine-cells and Transit-cells. The Engine-cell is differentiable for architecture search, while the Transit-cell only transits a sub-graph by architecture derivation. Consequently, the gap between the architectures in search and evaluation is significantly reduced. Our method also spares much memory and computation cost, which speeds up the search process. A feature sharing strategy is introduced for more balanced optimization and more efficient search. Furthermore, we develop an architecture derivation method to replace the traditional one that is based on a hand-crafted rule. Our method enables differentiable sparsification, and keeps the derived architecture equivalent to that of Engine-cell, which further improves the consistency between search and evaluation. Besides, it supports the search for topology where a node can be connected to prior nodes with any number of connections, so that the searched architectures could be more flexible. For experiments on CIFAR-10, our search on the standard space requires only 0.06 GPU-day. We further have an error rate of 2.22% with 0.07 GPU-day for the search on an extended space. We can also directly perform the search on ImageNet with topology learnable and achieve a top-1 error rate of 23.8% in 2.1 GPU-day.

Tao Huang, Shan You, Yibo Yang et al.

Differentiable neural architecture search (DARTS) has gained much success in discovering flexible and diverse cell types. To reduce the evaluation gap, the supernet is expected to have identical layers with the target network. However, even for this consistent search, the searched cells often suffer from poor performance, especially for the supernet with fewer layers, as current DARTS methods are prone to wide and shallow cells, and this topology collapse induces sub-optimal searched cells. In this paper, we alleviate this issue by endowing the cells with explicit stretchability, so the search can be directly implemented on our stretchable cells for both operation type and topology simultaneously. Concretely, we introduce a set of topological variables and a combinatorial probabilistic distribution to explicitly model the target topology. With more diverse and complex topologies, our method adapts well for various layer numbers. Extensive experiments on CIFAR-10 and ImageNet show that our stretchable cells obtain better performance with fewer layers and parameters. For example, our method can improve DARTS by 0.28\% accuracy on CIFAR-10 dataset with 45\% parameters reduced or 2.9\% with similar FLOPs on ImageNet dataset.

Ruihan Yang, Yibo Yang, Joseph Marino et al.

Recent work by Marino et al. (2020) showed improved performance in sequential density estimation by combining masked autoregressive flows with hierarchical latent variable models. We draw a connection between such autoregressive generative models and the task of lossy video compression. Specifically, we view recent neural video compression methods (Lu et al., 2019; Yang et al., 2020b; Agustssonet al., 2020) as instances of a generalized stochastic temporal autoregressive transform, and propose avenues for enhancement based on this insight. Comprehensive evaluations on large-scale video data show improved rate-distortion performance over both state-of-the-art neural and conventional video compression methods.

Yibo Yang, Hongyang Li, Shan You et al.

Neural architecture search (NAS) aims to produce the optimal sparse solution from a high-dimensional space spanned by all candidate connections. Current gradient-based NAS methods commonly ignore the constraint of sparsity in the search phase, but project the optimized solution onto a sparse one by post-processing. As a result, the dense super-net for search is inefficient to train and has a gap with the projected architecture for evaluation. In this paper, we formulate neural architecture search as a sparse coding problem. We perform the differentiable search on a compressed lower-dimensional space that has the same validation loss as the original sparse solution space, and recover an architecture by solving the sparse coding problem. The differentiable search and architecture recovery are optimized in an alternate manner. By doing so, our network for search at each update satisfies the sparsity constraint and is efficient to train. In order to also eliminate the depth and width gap between the network in search and the target-net in evaluation, we further propose a method to search and evaluate in one stage under the target-net settings. When training finishes, architecture variables are absorbed into network weights. Thus we get the searched architecture and optimized parameters in a single run. In experiments, our two-stage method on CIFAR-10 requires only 0.05 GPU-day for search. Our one-stage method produces state-of-the-art performances on both CIFAR-10 and ImageNet at the cost of only evaluation time.

Yibo Yang, Robert Bamler, Stephan Mandt

We consider the problem of lossy image compression with deep latent variable models. State-of-the-art methods build on hierarchical variational autoencoders (VAEs) and learn inference networks to predict a compressible latent representation of each data point. Drawing on the variational inference perspective on compression, we identify three approximation gaps which limit performance in the conventional approach: an amortization gap, a discretization gap, and a marginalization gap. We propose remedies for each of these three limitations based on ideas related to iterative inference, stochastic annealing for discrete optimization, and bits-back coding, resulting in the first application of bits-back coding to lossy compression. In our experiments, which include extensive baseline comparisons and ablation studies, we achieve new state-of-the-art performance on lossy image compression using an established VAE architecture, by changing only the inference method.

Xia Li, Yibo Yang, Qijie Zhao et al.

The convolution operation suffers from a limited receptive filed, while global modeling is fundamental to dense prediction tasks, such as semantic segmentation. In this paper, we apply graph convolution into the semantic segmentation task and propose an improved Laplacian. The graph reasoning is directly performed in the original feature space organized as a spatial pyramid. Different from existing methods, our Laplacian is data-dependent and we introduce an attention diagonal matrix to learn a better distance metric. It gets rid of projecting and re-projecting processes, which makes our proposed method a light-weight module that can be easily plugged into current computer vision architectures. More importantly, performing graph reasoning directly in the feature space retains spatial relationships and makes spatial pyramid possible to explore multiple long-range contextual patterns from different scales. Experiments on Cityscapes, COCO Stuff, PASCAL Context and PASCAL VOC demonstrate the effectiveness of our proposed methods on semantic segmentation. We achieve comparable performance with advantages in computational and memory overhead.

Yibo Yang, Robert Bamler, Stephan Mandt

We propose a novel algorithm for quantizing continuous latent representations in trained models. Our approach applies to deep probabilistic models, such as variational autoencoders (VAEs), and enables both data and model compression. Unlike current end-to-end neural compression methods that cater the model to a fixed quantization scheme, our algorithm separates model design and training from quantization. Consequently, our algorithm enables "plug-and-play" compression with variable rate-distortion trade-off, using a single trained model. Our algorithm can be seen as a novel extension of arithmetic coding to the continuous domain, and uses adaptive quantization accuracy based on estimates of posterior uncertainty. Our experimental results demonstrate the importance of taking into account posterior uncertainties, and show that image compression with the proposed algorithm outperforms JPEG over a wide range of bit rates using only a single standard VAE. Further experiments on Bayesian neural word embeddings demonstrate the versatility of the proposed method.

Yibo Yang, Hongyang Li, Xia Li et al.

The panoptic segmentation task requires a unified result from semantic and instance segmentation outputs that may contain overlaps. However, current studies widely ignore modeling overlaps. In this study, we aim to model overlap relations among instances and resolve them for panoptic segmentation. Inspired by scene graph representation, we formulate the overlapping problem as a simplified case, named scene overlap graph. We leverage each object's category, geometry and appearance features to perform relational embedding, and output a relation matrix that encodes overlap relations. In order to overcome the lack of supervision, we introduce a differentiable module to resolve the overlap between any pair of instances. The mask logits after removing overlaps are fed into per-pixel instance \verb|id| classification, which leverages the panoptic supervision to assist in the modeling of overlap relations. Besides, we generate an approximate ground truth of overlap relations as the weak supervision, to quantify the accuracy of overlap relations predicted by our method. Experiments on COCO and Cityscapes demonstrate that our method is able to accurately predict overlap relations, and outperform the state-of-the-art performance for panoptic segmentation. Our method also won the Innovation Award in COCO 2019 challenge.

Xia Li, Zhisheng Zhong, Jianlong Wu et al.

Self-attention mechanism has been widely used for various tasks. It is designed to compute the representation of each position by a weighted sum of the features at all positions. Thus, it can capture long-range relations for computer vision tasks. However, it is computationally consuming. Since the attention maps are computed w.r.t all other positions. In this paper, we formulate the attention mechanism into an expectation-maximization manner and iteratively estimate a much more compact set of bases upon which the attention maps are computed. By a weighted summation upon these bases, the resulting representation is low-rank and deprecates noisy information from the input. The proposed Expectation-Maximization Attention (EMA) module is robust to the variance of input and is also friendly in memory and computation. Moreover, we set up the bases maintenance and normalization methods to stabilize its training procedure. We conduct extensive experiments on popular semantic segmentation benchmarks including PASCAL VOC, PASCAL Context and COCO Stuff, on which we set new records.

Huan Li, Yibo Yang, Dongmin Chen et al.

Deep neural networks have been one of the dominant machine learning approaches in recent years. Several new network structures are proposed and have better performance than the traditional feedforward neural network structure. Representative ones include the skip connection structure in ResNet and the dense connection structure in DenseNet. However, it still lacks a unified guidance for the neural network structure design. In this paper, we propose the hypothesis that the neural network structure design can be inspired by optimization algorithms and a faster optimization algorithm may lead to a better neural network structure. Specifically, we prove that the propagation in the feedforward neural network with the same linear transformation in different layers is equivalent to minimizing some function using the gradient descent algorithm. Based on this observation, we replace the gradient descent algorithm with the heavy ball algorithm and Nesterov's accelerated gradient descent algorithm, which are faster and inspire us to design new and better network structures. ResNet and DenseNet can be considered as two special cases of our framework. Numerical experiments on CIFAR-10, CIFAR-100 and ImageNet verify the advantage of our optimization algorithm inspired structures over ResNet and DenseNet.

Zhisheng Zhong, Tiancheng Shen, Yibo Yang et al.

Convolutional neural networks (CNNs) have recently achieved great success in single-image super-resolution (SISR). However, these methods tend to produce over-smoothed outputs and miss some textural details. To solve these problems, we propose the Super-Resolution CliqueNet (SRCliqueNet) to reconstruct the high resolution (HR) image with better textural details in the wavelet domain. The proposed SRCliqueNet firstly extracts a set of feature maps from the low resolution (LR) image by the clique blocks group. Then we send the set of feature maps to the clique up-sampling module to reconstruct the HR image. The clique up-sampling module consists of four sub-nets which predict the high resolution wavelet coefficients of four sub-bands. Since we consider the edge feature properties of four sub-bands, the four sub-nets are connected to the others so that they can learn the coefficients of four sub-bands jointly. Finally we apply inverse discrete wavelet transform (IDWT) to the output of four sub-nets at the end of the clique up-sampling module to increase the resolution and reconstruct the HR image. Extensive quantitative and qualitative experiments on benchmark datasets show that our method achieves superior performance over the state-of-the-art methods.

Yibo Yang, Nicholas Ruozzi, Vibhav Gogate

We propose a simple and easy to implement neural network compression algorithm that achieves results competitive with more complicated state-of-the-art methods. The key idea is to modify the original optimization problem by adding K independent Gaussian priors (corresponding to the k-means objective) over the network parameters to achieve parameter quantization, as well as an L1 penalty to achieve pruning. Unlike many existing quantization-based methods, our method uses hard clustering assignments of network parameters, which adds minimal change or overhead to standard network training. We also demonstrate experimentally that tying neural network parameters provides less gain in generalization performance than changing network architecture and connectivity patterns entirely.

Yibo Yang, Zhisheng Zhong, Tiancheng Shen et al.

Improving information flow in deep networks helps to ease the training difficulties and utilize parameters more efficiently. Here we propose a new convolutional neural network architecture with alternately updated clique (CliqueNet). In contrast to prior networks, there are both forward and backward connections between any two layers in the same block. The layers are constructed as a loop and are updated alternately. The CliqueNet has some unique properties. For each layer, it is both the input and output of any other layer in the same block, so that the information flow among layers is maximized. During propagation, the newly updated layers are concatenated to re-update previously updated layer, and parameters are reused for multiple times. This recurrent feedback structure is able to bring higher level visual information back to refine low-level filters and achieve spatial attention. We analyze the features generated at different stages and observe that using refined features leads to a better result. We adopt a multi-scale feature strategy that effectively avoids the progressive growth of parameters. Experiments on image recognition datasets including CIFAR-10, CIFAR-100, SVHN and ImageNet show that our proposed models achieve the state-of-the-art performance with fewer parameters.