Search

Vision foundation models are renowned for their generalization ability due to massive training data. Nevertheless, they demand tremendous training resources, and the training data is often inaccessible, e.g., CLIP, DINOv2, posing great challenges to developing derivatives that could advance research in this field. In this work, we offer a very simple and general solution, named Proteus, to distill foundation models into smaller equivalents on ImageNet-1K without access to the original training data. Specifically, we remove the designs from conventional knowledge distillation settings that result in dataset bias and present three levels of training objectives, i.e., token, patch, and feature, to maximize the efficacy of knowledge transfer. In this manner, Proteus is trained at ImageNet-level costs with surprising ability, facilitating the accessibility of training foundation models for the broader research community. Leveraging DINOv2-g/14 as the teacher, Proteus-L/14 matches the performance of the Oracle method DINOv2-L/14 (142M training data) across 15 benchmarks and outperforms other vision foundation models including CLIP-L/14 (400M), OpenCLIP-L/14 (400M/2B) and SynCLR-L/14 (600M).

Training large language models (LLMs) and multimodal LLMs necessitates significant computing resources, and existing publicly available LLMs are typically pre-trained on diverse, privately curated datasets spanning various tasks. For instance, LLaMA, Vicuna, and LLaVA are three LLM variants trained with LLaMA base models using very different training recipes, tasks, and data modalities. The training cost and complexity for such LLM variants grow rapidly. In this study, we propose to use a soup strategy to assemble these LLM variants into a single well-generalized multimodal LLM (SoupLM) in a cost-efficient manner. Assembling these LLM variants efficiently brings knowledge and specialities trained from different domains and data modalities into an integrated one (e.g., chatbot speciality from user-shared conversations for Vicuna, and visual capacity from vision-language data for LLaVA), therefore, to avoid computing costs of repetitive training on several different domains. We propose series of soup strategies to systematically benchmark performance gains across various configurations, and probe the soup behavior across base models in the interpolation space.

Recent studies have drawn attention to the untapped potential of the "star operation" (element-wise multiplication) in network design. While intuitive explanations abound, the foundational rationale behind its application remains largely unexplored. Our study attempts to reveal the star operation's ability to map inputs into high-dimensional, non-linear feature spaces -- akin to kernel tricks -- without widening the network. We further introduce StarNet, a simple yet powerful prototype, demonstrating impressive performance and low latency under compact network structure and efficient budget. Like stars in the sky, the star operation appears unremarkable but holds a vast universe of potential. Our work encourages further exploration across tasks, with codes available at https://github.com/ma-xu/Rewrite-the-Stars., Comment: Accepted by CVPR 2024. Codes are made publically available at https://github.com/ma-xu/Rewrite-the-Stars

Current training pipelines in object recognition neglect Hue Jittering when doing data augmentation as it not only brings appearance changes that are detrimental to classification, but also the implementation is inefficient in practice. In this study, we investigate the effect of hue variance in the context of video understanding and find this variance to be beneficial since static appearances are less important in videos that contain motion information. Based on this observation, we propose a data augmentation method for video understanding, named Motion Coherent Augmentation (MCA), that introduces appearance variation in videos and implicitly encourages the model to prioritize motion patterns, rather than static appearances. Concretely, we propose an operation SwapMix to efficiently modify the appearance of video samples, and introduce Variation Alignment (VA) to resolve the distribution shift caused by SwapMix, enforcing the model to learn appearance invariant representations. Comprehensive empirical evaluation across various architectures and different datasets solidly validates the effectiveness and generalization ability of MCA, and the application of VA in other augmentation methods. Code is available at https://github.com/BeSpontaneous/MCA-pytorch., Comment: Accepted by ICLR2024

Latent graph inference (LGI) aims to jointly learn the underlying graph structure and node representations from data features. However, existing LGI methods commonly suffer from the issue of supervision starvation, where massive edge weights are learned without semantic supervision and do not contribute to the training loss. Consequently, these supervision-starved weights, which may determine the predictions of testing samples, cannot be semantically optimal, resulting in poor generalization. In this paper, we observe that this issue is actually caused by the graph sparsification operation, which severely destroys the important connections established between pivotal nodes and labeled ones. To address this, we propose to restore the corrupted affinities and replenish the missed supervision for better LGI. The key challenge then lies in identifying the critical nodes and recovering the corrupted affinities. We begin by defining the pivotal nodes as $k$-hop starved nodes, which can be identified based on a given adjacency matrix. Considering the high computational burden, we further present a more efficient alternative inspired by CUR matrix decomposition. Subsequently, we eliminate the starved nodes by reconstructing the destroyed connections. Extensive experiments on representative benchmarks demonstrate that reducing the starved nodes consistently improves the performance of state-of-the-art LGI methods, especially under extremely limited supervision (6.12% improvement on Pubmed with a labeling rate of only 0.3%).

Existing video recognition algorithms always conduct different training pipelines for inputs with different frame numbers, which requires repetitive training operations and multiplying storage costs. If we evaluate the model using other frames which are not used in training, we observe the performance will drop significantly (see Fig.1), which is summarized as Temporal Frequency Deviation phenomenon. To fix this issue, we propose a general framework, named Frame Flexible Network (FFN), which not only enables the model to be evaluated at different frames to adjust its computation, but also reduces the memory costs of storing multiple models significantly. Concretely, FFN integrates several sets of training sequences, involves Multi-Frequency Alignment (MFAL) to learn temporal frequency invariant representations, and leverages Multi-Frequency Adaptation (MFAD) to further strengthen the representation abilities. Comprehensive empirical validations using various architectures and popular benchmarks solidly demonstrate the effectiveness and generalization of FFN (e.g., 7.08/5.15/2.17% performance gain at Frame 4/8/16 on Something-Something V1 dataset over Uniformer). Code is available at https://github.com/BeSpontaneous/FFN., Comment: Accepted by CVPR2023

Anomaly detection in videos is a significant yet challenging problem. Previous approaches based on deep neural networks employ either reconstruction-based or prediction-based approaches. Nevertheless, existing reconstruction-based methods 1) rely on old-fashioned convolutional autoencoders and are poor at modeling temporal dependency; 2) are prone to overfit the training samples, leading to indistinguishable reconstruction errors of normal and abnormal frames during the inference phase. To address such issues, firstly, we get inspiration from transformer and propose ${\textbf S}$patio-${\textbf T}$emporal ${\textbf A}$uto-${\textbf T}$rans-${\textbf E}$ncoder, dubbed as $\textbf{STATE}$, as a new autoencoder model for enhanced consecutive frame reconstruction. Our STATE is equipped with a specifically designed learnable convolutional attention module for efficient temporal learning and reasoning. Secondly, we put forward a novel reconstruction-based input perturbation technique during testing to further differentiate anomalous frames. With the same perturbation magnitude, the testing reconstruction error of the normal frames lowers more than that of the abnormal frames, which contributes to mitigating the overfitting problem of reconstruction. Owing to the high relevance of the frame abnormality and the objects in the frame, we conduct object-level reconstruction using both the raw frame and the corresponding optical flow patches. Finally, the anomaly score is designed based on the combination of the raw and motion reconstruction errors using perturbed inputs. Extensive experiments on benchmark video anomaly detection datasets demonstrate that our approach outperforms previous reconstruction-based methods by a notable margin, and achieves state-of-the-art anomaly detection performance consistently. The code is available at https://github.com/wyzjack/MRMGA4VAD., Comment: Accepted by ICDM 2022

The state of neural network pruning has been noticed to be unclear and even confusing for a while, largely due to "a lack of standardized benchmarks and metrics" [3]. To standardize benchmarks, first, we need to answer: what kind of comparison setup is considered fair? This basic yet crucial question has barely been clarified in the community, unfortunately. Meanwhile, we observe several papers have used (severely) sub-optimal hyper-parameters in pruning experiments, while the reason behind them is also elusive. These sub-optimal hyper-parameters further exacerbate the distorted benchmarks, rendering the state of neural network pruning even more obscure. Two mysteries in pruning represent such a confusing status: the performance-boosting effect of a larger finetuning learning rate, and the no-value argument of inheriting pretrained weights in filter pruning. In this work, we attempt to explain the confusing state of network pruning by demystifying the two mysteries. Specifically, (1) we first clarify the fairness principle in pruning experiments and summarize the widely-used comparison setups; (2) then we unveil the two pruning mysteries and point out the central role of network trainability, which has not been well recognized so far; (3) finally, we conclude the paper and give some concrete suggestions regarding how to calibrate the pruning benchmarks in the future. Code: https://github.com/mingsun-tse/why-the-state-of-pruning-so-confusing., Comment: 17 pages, v2, corrected wrong references

Existing action recognition methods typically sample a few frames to represent each video to avoid the enormous computation, which often limits the recognition performance. To tackle this problem, we propose Ample and Focal Network (AFNet), which is composed of two branches to utilize more frames but with less computation. Specifically, the Ample Branch takes all input frames to obtain abundant information with condensed computation and provides the guidance for Focal Branch by the proposed Navigation Module; the Focal Branch squeezes the temporal size to only focus on the salient frames at each convolution block; in the end, the results of two branches are adaptively fused to prevent the loss of information. With this design, we can introduce more frames to the network but cost less computation. Besides, we demonstrate AFNet can utilize fewer frames while achieving higher accuracy as the dynamic selection in intermediate features enforces implicit temporal modeling. Further, we show that our method can be extended to reduce spatial redundancy with even less cost. Extensive experiments on five datasets demonstrate the effectiveness and efficiency of our method., Comment: Accepted by NeurIPS 2022

A deeper network structure generally handles more complicated non-linearity and performs more competitively. Nowadays, advanced network designs often contain a large number of repetitive structures (e.g., Transformer). They empower the network capacity to a new level but also increase the model size inevitably, which is unfriendly to either model restoring or transferring. In this study, we are the first to investigate the representative potential of fixed random weights with limited unique values by learning diverse masks and introduce the Parameter-Efficient Masking Networks (PEMN). It also naturally leads to a new paradigm for model compression to diminish the model size. Concretely, motivated by the repetitive structures in modern neural networks, we utilize one random initialized layer, accompanied with different masks, to convey different feature mappings and represent repetitive network modules. Therefore, the model can be expressed as \textit{one-layer} with a bunch of masks, which significantly reduce the model storage cost. Furthermore, we enhance our strategy by learning masks for a model filled by padding a given random weights vector. In this way, our method can further lower the space complexity, especially for models without many repetitive architectures. We validate the potential of PEMN learning masks on random weights with limited unique values and test its effectiveness for a new compression paradigm based on different network architectures. Code is available at https://github.com/yueb17/PEMN

In recent years, change point detection for high dimensional data has become increasingly important in many scientific fields. Most literature develop a variety of separate methods designed for specified models (e.g. mean shift model, vector auto-regressive model, graphical model). In this paper, we provide a unified framework for structural break detection which is suitable for a large class of models. Moreover, the proposed algorithm automatically achieves consistent parameter estimates during the change point detection process, without the need for refitting the model. Specifically, we introduce a three-step procedure. The first step utilizes the block segmentation strategy combined with a fused lasso based estimation criterion, leads to significant computational gains without compromising the statistical accuracy in identifying the number and location of the structural breaks. This procedure is further coupled with hard-thresholding and exhaustive search steps to consistently estimate the number and location of the break points. The strong guarantees are proved on both the number of estimated change points and the rates of convergence of their locations. The consistent estimates of model parameters are also provided. The numerical studies provide further support of the theory and validate its competitive performance for a wide range of models. The developed algorithm is implemented in the R package LinearDetect.

Fully exploiting the learning capacity of neural networks requires overparameterized dense networks. On the other side, directly training sparse neural networks typically results in unsatisfactory performance. Lottery Ticket Hypothesis (LTH) provides a novel view to investigate sparse network training and maintain its capacity. Concretely, it claims there exist winning tickets from a randomly initialized network found by iterative magnitude pruning and preserving promising trainability (or we say being in trainable condition). In this work, we regard the winning ticket from LTH as the subnetwork which is in trainable condition and its performance as our benchmark, then go from a complementary direction to articulate the Dual Lottery Ticket Hypothesis (DLTH): Randomly selected subnetworks from a randomly initialized dense network can be transformed into a trainable condition and achieve admirable performance compared with LTH -- random tickets in a given lottery pool can be transformed into winning tickets. Specifically, by using uniform-randomly selected subnetworks to represent the general cases, we propose a simple sparse network training strategy, Random Sparse Network Transformation (RST), to substantiate our DLTH. Concretely, we introduce a regularization term to borrow learning capacity and realize information extrusion from the weights which will be masked. After finishing the transformation for the randomly selected subnetworks, we conduct the regular finetuning to evaluate the model using fair comparisons with LTH and other strong baselines. Extensive experiments on several public datasets and comparisons with competitive approaches validate our DLTH as well as the effectiveness of the proposed model RST. Our work is expected to pave a way for inspiring new research directions of sparse network training in the future. Our code is available at https://github.com/yueb17/DLTH.

The nanoscale structure of molecular assemblies plays a major role in many ($\mu$)-biological mechanisms. Molecular crystals are one of the most simple of these assemblies and are widely used in a variety of applications from pharmaceuticals and agrochemicals, to nutraceuticals and cosmetics. The collective vibrations in such molecular crystals can be probed using terahertz spectroscopy, providing unique characteristic spectral fingerprints. However, the association of the spectral features to the crystal conformation, crystal phase and its environment is a difficult task. We present a combined computationalexperimental study on the incorporation of water in lactose molecular crystals, and show how simulations can be used to associate spectral features in the THz region to crystal conformations and phases. Using periodic DFT simulations of lactose molecular crystals, the role of water in the observed lactose THz spectrum is clarified, presenting both direct and indirect contributions. A specific experimental setup is built to allow the controlled heating and corresponding dehydration of the sample, providing the monitoring of the crystal phase transformation dynamics. Besides the observation that lactose phases and phase transformation appear to be more complex than previously thoughtincluding several crystal forms in a single phase and a non-negligible water content in the so-called anhydrous phasewe draw two main conclusions from this study. Firstly, THz modes are spread over more than one molecule and require periodic computation rather than a gas-phase one. Secondly, hydration water does not only play a perturbative role but also participates in the facilitation of the THz vibrations.

The fast transmission rate of COVID-19 worldwide has made this virus the most important challenge of year 2020. Many mitigation policies have been imposed by the governments at different regional levels (country, state, county, and city) to stop the spread of this virus. Quantifying the effect of such mitigation strategies on the transmission and recovery rates, and predicting the rate of new daily cases are two crucial tasks. In this paper, we propose a hybrid modeling framework which not only accounts for such policies but also utilizes the spatial and temporal information to characterize the pattern of COVID-19 progression. Specifically, a piecewise susceptible-infected-recovered (SIR) model is developed while the dates at which the transmission/recover rates change significantly are defined as "break points" in this model. A novel and data-driven algorithm is designed to locate the break points using ideas from fused lasso and thresholding. In order to enhance the forecasting power and to describe additional temporal dependence among the daily number of cases, this model is further coupled with spatial smoothing covariates and vector auto-regressive (VAR) model. The proposed model is applied to several U.S. states and counties, and the results confirm the effect of "stay-at-home orders" and some states' early "re-openings" by detecting break points close to such events. Further, the model provided satisfactory short-term forecasts of the number of new daily cases at regional levels by utilizing the estimated spatio-temporal covariance structures. They were also better or on par with other proposed models in the literature, including flexible deep learning ones. Finally, selected theoretical results and empirical performance of the proposed methodology on synthetic data are reported which justify the good performance of the proposed method.

Action prediction aims to infer the forthcoming human action with partially-observed videos, which is a challenging task due to the limited information underlying early observations. Existing methods mainly adopt a reconstruction strategy to handle this task, expecting to learn a single mapping function from partial observations to full videos to facilitate the prediction process. In this study, we propose adversarial memory networks (AMemNet) to generate the "full video" feature conditioning on a partial video query from two new aspects. Firstly, a key-value structured memory generator is designed to memorize different partial videos as key memories and dynamically write full videos in value memories with gating mechanism and querying attention. Secondly, we develop a class-aware discriminator to guide the memory generator to deliver not only realistic but also discriminative full video features upon adversarial training. The final prediction result of AMemNet is given by late fusion over RGB and optical flow streams. Extensive experimental results on two benchmark video datasets, UCF-101 and HMDB51, are provided to demonstrate the effectiveness of the proposed AMemNet model over state-of-the-art methods.

Anomaly detection is a fundamental yet challenging problem in machine learning due to the lack of label information. In this work, we propose a novel and powerful framework, dubbed as SLA$^2$P, for unsupervised anomaly detection. After extracting representative embeddings from raw data, we apply random projections to the features and regard features transformed by different projections as belonging to distinct pseudo classes. We then train a classifier network on these transformed features to perform self-supervised learning. Next we add adversarial perturbation to the transformed features to decrease their softmax scores of the predicted labels and design anomaly scores based on the predictive uncertainties of the classifier on these perturbed features. Our motivation is that because of the relatively small number and the decentralized modes of anomalies, 1) the pseudo label classifier's training concentrates more on learning the semantic information of normal data rather than anomalous data; 2) the transformed features of the normal data are more robust to the perturbations than those of the anomalies. Consequently, the perturbed transformed features of anomalies fail to be classified well and accordingly have lower anomaly scores than those of the normal samples. Extensive experiments on image, text and inherently tabular benchmark datasets back up our findings and indicate that SLA$^2$P achieves state-of-the-art results on unsupervised anomaly detection tasks consistently.

Sign language is commonly used by deaf or mute people to communicate but requires extensive effort to master. It is usually performed with the fast yet delicate movement of hand gestures, body posture, and even facial expressions. Current Sign Language Recognition (SLR) methods usually extract features via deep neural networks and suffer overfitting due to limited and noisy data. Recently, skeleton-based action recognition has attracted increasing attention due to its subject-invariant and background-invariant nature, whereas skeleton-based SLR is still under exploration due to the lack of hand annotations. Some researchers have tried to use off-line hand pose trackers to obtain hand keypoints and aid in recognizing sign language via recurrent neural networks. Nevertheless, none of them outperforms RGB-based approaches yet. To this end, we propose a novel Skeleton Aware Multi-modal Framework with a Global Ensemble Model (GEM) for isolated SLR (SAM-SLR-v2) to learn and fuse multi-modal feature representations towards a higher recognition rate. Specifically, we propose a Sign Language Graph Convolution Network (SL-GCN) to model the embedded dynamics of skeleton keypoints and a Separable Spatial-Temporal Convolution Network (SSTCN) to exploit skeleton features. The skeleton-based predictions are fused with other RGB and depth based modalities by the proposed late-fusion GEM to provide global information and make a faithful SLR prediction. Experiments on three isolated SLR datasets demonstrate that our proposed SAM-SLR-v2 framework is exceedingly effective and achieves state-of-the-art performance with significant margins. Our code will be available at https://github.com/jackyjsy/SAM-SLR-v2

Vector Auto-Regressive (VAR) models capture lead-lag temporal dynamics of multivariate time series data. They have been widely used in macroeconomics, financial econometrics, neuroscience and functional genomics. In many applications, the data exhibit structural changes in their autoregressive dynamics, which correspond to changes in the transition matrices of the VAR model that specify such dynamics. We present the R package VARDetect that implements two classes of algorithms to detect multiple change points in piecewise stationary VAR models. The first exhibits sublinear computational complexity in the number of time points and is best suited for structured sparse models, while the second exhibits linear time complexity and is designed for models whose transition matrices are assumed to have a low rank plus sparse decomposition. The package also has functions to generate data from the various variants of VAR models discussed, which is useful in simulation studies, as well as to visualize the results through network layouts.

Several recent works [40, 24] observed an interesting phenomenon in neural network pruning: A larger finetuning learning rate can improve the final performance significantly. Unfortunately, the reason behind it remains elusive up to date. This paper is meant to explain it through the lens of dynamical isometry [42]. Specifically, we examine neural network pruning from an unusual perspective: pruning as initialization for finetuning, and ask whether the inherited weights serve as a good initialization for the finetuning? The insights from dynamical isometry suggest a negative answer. Despite its critical role, this issue has not been well-recognized by the community so far. In this paper, we will show the understanding of this problem is very important -- on top of explaining the aforementioned mystery about the larger finetuning rate, it also unveils the mystery about the value of pruning [5, 30]. Besides a clearer theoretical understanding of pruning, resolving the problem can also bring us considerable performance benefits in practice., Comment: 8 pages, 2 figures, 7 tables

Sign language is commonly used by deaf or speech impaired people to communicate but requires significant effort to master. Sign Language Recognition (SLR) aims to bridge the gap between sign language users and others by recognizing signs from given videos. It is an essential yet challenging task since sign language is performed with the fast and complex movement of hand gestures, body posture, and even facial expressions. Recently, skeleton-based action recognition attracts increasing attention due to the independence between the subject and background variation. However, skeleton-based SLR is still under exploration due to the lack of annotations on hand keypoints. Some efforts have been made to use hand detectors with pose estimators to extract hand key points and learn to recognize sign language via Neural Networks, but none of them outperforms RGB-based methods. To this end, we propose a novel Skeleton Aware Multi-modal SLR framework (SAM-SLR) to take advantage of multi-modal information towards a higher recognition rate. Specifically, we propose a Sign Language Graph Convolution Network (SL-GCN) to model the embedded dynamics and a novel Separable Spatial-Temporal Convolution Network (SSTCN) to exploit skeleton features. RGB and depth modalities are also incorporated and assembled into our framework to provide global information that is complementary to the skeleton-based methods SL-GCN and SSTCN. As a result, SAM-SLR achieves the highest performance in both RGB (98.42\%) and RGB-D (98.53\%) tracks in 2021 Looking at People Large Scale Signer Independent Isolated SLR Challenge. Our code is available at https://github.com/jackyjsy/CVPR21Chal-SLR, Comment: This is a preprint version of our work SAM-SLR that ranked 1st at CVPR2021 Challenge on Large Scale Signer Independent Isolated Sign Language Recognition

Neural network pruning typically removes connections or neurons from a pretrained converged model; while a new pruning paradigm, pruning at initialization (PaI), attempts to prune a randomly initialized network. This paper offers the first survey concentrated on this emerging pruning fashion. We first introduce a generic formulation of neural network pruning, followed by the major classic pruning topics. Then, as the main body of this paper, a thorough and structured literature review of PaI methods is presented, consisting of two major tracks (sparse training and sparse selection). Finally, we summarize the surge of PaI compared to PaT and discuss the open problems. Apart from the dedicated literature review, this paper also offers a code base for easy sanity-checking and benchmarking of different PaI methods., Comment: Accepted in IJCAI'22 Survey Track. Code base: https://github.com/mingsun-tse/smile-pruning

Advances in face rotation, along with other face-based generative tasks, are more frequent as we advance further in topics of deep learning. Even as impressive milestones are achieved in synthesizing faces, the importance of preserving identity is needed in practice and should not be overlooked. Also, the difficulty should not be more for data with obscured faces, heavier poses, and lower quality. Existing methods tend to focus on samples with variation in pose, but with the assumption data is high in quality. We propose a generative adversarial network (GAN) -based model to generate high-quality, identity preserving frontal faces from one or multiple low-resolution (LR) faces with extreme poses. Specifically, we propose SuperFront-GAN (SF-GAN) to synthesize a high-resolution (HR), frontal face from one-to-many LR faces with various poses and with the identity-preserved. We integrate a super-resolution (SR) side-view module into SF-GAN to preserve identity information and fine details of the side-views in HR space, which helps model reconstruct high-frequency information of faces (i.e., periocular, nose, and mouth regions). Moreover, SF-GAN accepts multiple LR faces as input, and improves each added sample. We squeeze additional gain in performance with an orthogonal constraint in the generator to penalize redundant latent representations and, hence, diversify the learned features space. Quantitative and qualitative results demonstrate the superiority of SF-GAN over others.

Multi-view action recognition (MVAR) leverages complementary temporal information from different views to improve the learning performance. Obtaining informative view-specific representation plays an essential role in MVAR. Attention has been widely adopted as an effective strategy for discovering discriminative cues underlying temporal data. However, most existing MVAR methods only utilize attention to extract representation for each view individually, ignoring the potential to dig latent patterns based on mutual-support information in attention space. To this end, we propose a collaborative attention mechanism (CAM) for solving the MVAR problem in this paper. The proposed CAM detects the attention differences among multi-view, and adaptively integrates frame-level information to benefit each other. Specifically, we extend the long short-term memory (LSTM) to a Mutual-Aid RNN (MAR) to achieve the multi-view collaboration process. CAM takes advantages of view-specific attention pattern to guide another view and discover potential information which is hard to be explored by itself. It paves a novel way to leverage attention information and enhances the multi-view representation learning. Extensive experiments on four action datasets illustrate the proposed CAM achieves better results for each view and also boosts multi-view performance.

Multi-view time series classification (MVTSC) aims to improve the performance by fusing the distinctive temporal information from multiple views. Existing methods mainly focus on fusing multi-view information at an early stage, e.g., by learning a common feature subspace among multiple views. However, these early fusion methods may not fully exploit the unique temporal patterns of each view in complicated time series. Moreover, the label correlations of multiple views, which are critical to boost-ing, are usually under-explored for the MVTSC problem. To address the aforementioned issues, we propose a Correlative Channel-Aware Fusion (C2AF) network. First, C2AF extracts comprehensive and robust temporal patterns by a two-stream structured encoder for each view, and captures the intra-view and inter-view label correlations with a graph-based correlation matrix. Second, a channel-aware learnable fusion mechanism is implemented through convolutional neural networks to further explore the global correlative patterns. These two steps are trained end-to-end in the proposed C2AF network. Extensive experimental results on three real-world datasets demonstrate the superiority of our approach over the state-of-the-art methods. A detailed ablation study is also provided to show the effectiveness of each model component.