Knowledge Distillation (KD) is a widely popular model compression technique that can effectively transfer knowledge from a pre-trained, large-scale teacher model to a more compact and lightweight student model. Traditional KD methods aim to improve the student’s representation capability by mimicking the teacher’s features, e.g., minimizing the \(\mathcal{L}_{2}\) distance between their intermediate features. However, due to the capacity gap between the student and the teacher, student often struggles to precisely mimic the features of the teacher. To address this challenge, we propose to boost the knowledge distillation for the visual recognition tasks via Direction Alignment and Discrepancy Standardization ( DADS) , which exploits the feature scaling technique to distill from both the feature direction and feature discrepancy. To this end, we devise an efficient feature alignment module to align the dimensions of teacher and student features. Moreover, we align the direction of student features and teacher features, which are pre-processed by normalization. Furthermore, we leverage the Kullback–Leibler (KL) divergence to refine the features alignment, minimizing discrepancy in the distribution of features across samples, which is pre-processed by \(\mathcal{Z}\) -score standardization. In this way, our proposed approach can effectively transfer the knowledge from the teacher to the student, facilitating the downstream visual recognition applications, such as image classification and semantic segmentation. Extensive experimental analyses clearly validate the effectiveness of DADS . Compared with previous KD methods, our approach sets a new benchmark, achieving state-of-the-art results on visual recognition tasks.

Recently, the bottleneck of multi-object tracking is shifting from detection performance to association performance. However, research on association algorithms requires a large number of identity labels, which are more expensive than detection labels. To circumvent the need for identity labels, we propose a Self-supervised Re-identification module for online joint Multi-Object Tracking (SR-MOT). Specifically, we design an appearance discriminator to judge identities based solely on detection hypotheses and then associate the same identity with the final trajectory. To train the discriminator without using identity labels, we construct negative pairs by the detections that appear in the same video frame, as they definitely belong to different identities. Positive pairs are naturally constructed through several useful data augmentation strategies at the box level. In addition, our proposed method balances conflicting detection and re-ID tasks by using different output features and dynamically adjusts detection and re-ID loss weights based on the information content of the loss distribution to promote balance between the two tasks from the feature level and optimization methods. In our evaluation on the MOT Challenge benchmark, we show that our SR-MOT performs comparably to supervised methods and is significantly superior to other unsupervised methods. Our proposed method provides a practical solution for multi-object tracking without the need for identity labels, making it more accessible for real-world applications.

Driven by empirical risk minimization, machine learning algorithm tends to exploit subtle statistical correlations existing in the training environment for prediction, while the spurious correlations are unstable across environments, leading to poor generalization performance. Accordingly, the problem of the Out-of-distribution (OOD) generalization aims to exploit an invariant/stable relationship between features and outcomes that generalizes well on all possible environments. To address the spurious correlation induced by the selection bias, in this article, we propose a novel Clique-based Causal Feature Separation (CCFS) algorithm by explicitly incorporating the causal structure to identify causal features of outcome for OOD generalization. Specifically, the proposed CCFS algorithm identifies the largest clique in the learned causal skeleton. Theoretically, we guarantee that either the largest clique or the rest of the causal skeleton is exactly the set of all causal features of the outcome. Finally, we separate the causal features from the non-causal ones with a sample-reweighting decorrelator for OOD prediction. Extensive experiments validate the effectiveness of the proposed CCFS method on both causal feature identification and OOD generalization tasks.

Deep clustering usually treats the clustering assignments as supervisory signals to learn a more compact representation with deep neural networks, under the guidance of clustering-oriented losses. Nevertheless, we observe that, without reliable supervision, such losses for global clustering would destroy the locally geometric structure underlying data. In this paper, we propose a local-to-global deep clustering method based on approximate uniform manifold (LGC-AUM) to address this issue in a two-stage fashion. In the local stage, an intra-manifold preservation loss is proposed to preserve intra-manifold structures locally on basis of approximate uniform manifold, and an inter-manifold discrimination loss is for global inter-manifold structure. Thus, this stage serves to learn more discriminative structure-preserving features by reducing the correlations between different manifolds, which paves the way for the final clustering. Build off the learned features, the second stage explores a clustering loss based on approximate uniform manifold to establish stable network training for effective clustering with two auxiliary distributions. Experiments on five benchmark datasets verify the efficacy of our LGC-AUM as compared to several well-behaved clustering counterparts.

Through using only a well-trained classifier, model-inversion (MI) attacks can recover the data used for training the classifier, leading to the privacy leakage of the training data. To defend against MI attacks, previous work utilizes a unilateral dependency optimization strategy, i.e., minimizing the dependency between inputs (i.e., features) and outputs (i.e., labels) during training the classifier. However, such a minimization process conflicts with minimizing the supervised loss that aims to maximize the dependency between inputs and outputs, causing an explicit trade-off between model robustness against MI attacks and model utility on classification tasks. In this paper, we aim to minimize the dependency between the latent representations and the inputs while maximizing the dependency between latent representations and the outputs, named a bilateral dependency optimization (BiDO) strategy. In particular, we use the dependency constraints as a universally applicable regularizer in addition to commonly used losses for deep neural networks (e.g., cross-entropy), which can be instantiated with appropriate dependency criteria according to different tasks. To verify the efficacy of our strategy, we propose two implementations of BiDO, by using two different dependency measures: BiDO with constrained covariance (BiDO-COCO) and BiDO with Hilbert-Schmidt Independence Criterion (BiDO-HSIC). Experiments show that BiDO achieves the state-of-the-art defense performance for a variety of datasets, classifiers, and MI attacks while suffering a minor classification-accuracy drop compared to the well-trained classifier with no defense, which lights up a novel road to defend against MI attacks.

Currently, many approaches to the sentence query based moment location (SQML) task emphasize (inter-)modality interaction between video and language query via transformer-based cross-attention or contrastive learning. However, they could still face two issues: 1) modality interaction could be unexpectedly friendly to modality specific learning that merely learns modality specific patterns, and 2) modality interaction easily confuses spatio-temporal cues and ultimately makes time cues in the original video ambiguous. In this paper, we propose a modality synergy with spatio-temporal cue purification method (MS2P) for SQML to address the above two issues. Particularly, a conceptually simple modality synergy strategy is explored to keep features modality specific while absorbing the other modality complementary information with both carefully designed cross-attention unit and non-contrastive learning. As a result, modality specific semantics can be calibrated progressively in a safer way. To preserve time cues in original video, we further purify video representation into spatial and temporal parts to enhance localization resolution by the proposed two light-weight sentence-aware filtering operations. Experiments on Charades-STA, TACoS, and ActivityNet Caption datasets show our model outperforms the state-of-the-art approaches by a large margin.

Semi-supervised learning (SSL) that utilizes plenty of unlabeled examples to boost the performance of learning from limited labeled examples is a powerful learning paradigm with widely real-world applications such as information retrieval and document clustering. Label propagation (LP) is a popular SSL method which propagates labels through the dataset along high density areas defined by unlabeled examples, but it is fragile to bridge examples. Semi-supervised K-Means uses labeled examples to initialize clustering centers to separate different examples, however, semi-supervised K-Means fails in the situation of imbalanced issues, that is, the example size of each class varies significantly. This paper proposes a novel label propagated nonnegative matrix factorization method (LPNMF) to handle clean labeled but biased data and its extension LPNMF-E to handle noisy labeled data based on the framework of NMF. LPNMF decomposes the whole dataset into the product of a basis matrix and a coefficient matrix. To propagate labels to unlabeled examples, LPNMF regards the class indicators of labeled examples as their coefficients and iteratively updates both basis matrix and coefficients of unlabeled examples. LPNMF absorbs the merits from both semi-supervised K-Means and label propagation to handle their respective shortages. Specifically, on the one hand, LPNMF learns representative clustering centers based on the distribution of the dataset, similar to semi-supervised K-means, and thus is robust to the bridge examples. On the other hand, LPNMF pushes labels according to the affinity between examples, similar to label propagation, and thus relieves the biased problem. Moreover, we introduce a LPNMF extension to handle the noisy label case. LPNMF-E relaxes the constraint of labeled examples. Since the label of each labeled example also obtains label information from the global distribution of the whole dataset and local manifold of its neighbors, LPNMF-E outputs reliable class indicators even if a portion of examples are incorrectly labeled. Theoretical analyses for the generalization ability of our proposed models are also provided. Experimental results on both clean and noisy labeled datasets confirm the effectiveness of LPNMF and LPNMF-E compared with both LP and the representative semi-supervised K-Means algorithms.

Most modern multi-object tracking (MOT) systems for videos follow the tracking-by-detection paradigm, where objects of interest are first located in each frame then associated correspondingly to form their intact trajectories. In this setting, the appearance features of objects usually provide the most important cues for data association, but it is very susceptible to occlusions, illumination variations, and inaccurate detections, thus easily resulting in incorrect trajectories. To address this issue, in this study we propose to make full use of the neighboring information. Our motivations derive from the observations that people tend to move in a group. As such, when an individual target's appearance is remarkably changed, the observer can still identify it with its neighbor context. To model the contextual information from neighbors, we first utilize the spatiotemporal relations among trajectories to efficiently select suitable neighbors for targets. Subsequently, we construct neighbor graph for each target and corresponding neighbors then employ the graph convolutional networks (GCNs) to model their relations and learn the graph features. To the best of our knowledge, it is the first time to explicitly leverage neighbor cues via GCN in MOT. Finally, standardized evaluations on the MOT16 and MOT17 data sets demonstrate that our approach can remarkably reduce the identity switches whilst achieve state-of-the-art overall performance.

Unsupervised domain adaptation (UDA) attempts to transfer specific knowledge from one domain with labeled data to another domain without labels. Recently, maximum squares loss has been proposed to tackle UDA problem but it does not consider the prediction diversity which has proven beneficial to UDA. In this paper, we propose a novel normalized squares maximization (NSM) loss in which the maximum squares is normalized by the sum of squares of class sizes. The normalization term enforces the class sizes of predictions to be balanced to explicitly increase the diversity. Theoretical analysis shows that the optimal solution to NSM is one-hot vectors with balanced class sizes, i.e., NSM encourages both discriminate and diverse predictions. We further propose a robust variant of NSM, RNSM, by replacing the square loss with L2,1-norm to reduce the influence of outliers and noises. Experiments of cross-domain image classification on two benchmark datasets illustrate the effectiveness of both NSM and RNSM. RNSM achieves promising performance compared to state-of-the-art methods. The code is available at https://github.com/wj-zhang/NSM.