Can Students Beyond The Teacher? Distilling Knowledge from Teacher's Bias

Knowledge distillation (KD) is a model compression technique that transfers knowledge from a large teacher model to a smaller student model to enhance its performance. Existing methods often assume that the student model is inherently inferior to the teacher model. However, we identify that the fundamental issue affecting student performance is the bias transferred by the teacher. Current KD frameworks transmit both right and wrong knowledge, introducing bias that misleads the student model. To address this issue, we propose a novel strategy to rectify bias and greatly improve the student model's performance. Our strategy involves three steps: First, we differentiate knowledge and design a bias elimination method to filter out biases, retaining only the right knowledge for the student model to learn. Next, we propose a bias rectification method to rectify the teacher model's wrong predictions, fundamentally addressing bias interference. The student model learns from both the right knowledge and the rectified biases, greatly improving its prediction accuracy. Additionally, we introduce a dynamic learning approach with a loss function that updates weights dynamically, allowing the student model to quickly learn right knowledge-based easy tasks initially and tackle hard tasks corresponding to biases later, greatly enhancing the student model's learning efficiency. To the best of our knowledge, this is the first strategy enabling the student model to surpass the teacher model. Experiments demonstrate that our strategy, as a plug-and-play module, is versatile across various mainstream KD frameworks. We will release our code after the paper is accepted.

Why pre-training is beneficial for downstream classification tasks?

Pre-training has exhibited notable benefits to downstream tasks by boosting accuracy and speeding up convergence, but the exact reasons for these benefits still remain unclear. To this end, we propose to quantitatively and explicitly explain effects of pre-training on the downstream task from a novel game-theoretic view, which also sheds new light into the learning behavior of deep neural networks (DNNs). Specifically, we extract and quantify the knowledge encoded by the pre-trained model, and further track the changes of such knowledge during the fine-tuning process. Interestingly, we discover that only a small amount of pre-trained model's knowledge is preserved for the inference of downstream tasks. However, such preserved knowledge is very challenging for a model training from scratch to learn. Thus, with the help of this exclusively learned and useful knowledge, the model fine-tuned from pre-training usually achieves better performance than the model training from scratch. Besides, we discover that pre-training can guide the fine-tuned model to learn target knowledge for the downstream task more directly and quickly, which accounts for the faster convergence of the fine-tuned model.

Layerwise Change of Knowledge in Neural Networks

This paper aims to explain how a deep neural network (DNN) gradually extracts new knowledge and forgets noisy features through layers in forward propagation. Up to now, although the definition of knowledge encoded by the DNN has not reached a consensus, Previous studies have derived a series of mathematical evidence to take interactions as symbolic primitive inference patterns encoded by a DNN. We extend the definition of interactions and, for the first time, extract interactions encoded by intermediate layers. We quantify and track the newly emerged interactions and the forgotten interactions in each layer during the forward propagation, which shed new light on the learning behavior of DNNs. The layer-wise change of interactions also reveals the change of the generalization capacity and instability of feature representations of a DNN.

Staircase Cascaded Fusion of Lightweight Local Pattern Recognition and Long-Range Dependencies for Structural Crack Segmentation

Detecting cracks with pixel-level precision for key structures is a significant challenge, as existing methods struggle to effectively integrate local textures and pixel dependencies of cracks. Furthermore, these methods often possess numerous parameters and substantial computational requirements, complicating deployment on edge devices. In this paper, we propose a staircase cascaded fusion crack segmentation network (CrackSCF) that generates high-quality crack segmentation maps using minimal computational resources. We constructed a staircase cascaded fusion module that effectively captures local patterns of cracks and long-range dependencies of pixels, and it can suppress background noise well. To reduce the computational resources required by the model, we introduced a lightweight convolution block, which replaces all convolution operations in the network, significantly reducing the required computation and parameters without affecting the network's performance. To evaluate our method, we created a challenging benchmark dataset called TUT and conducted experiments on this dataset and five other public datasets. The experimental results indicate that our method offers significant advantages over existing methods, especially in handling background noise interference and detailed crack segmentation. The F1 and mIoU scores on the TUT dataset are 0.8382 and 0.8473, respectively, achieving state-of-the-art (SOTA) performance while requiring the least computational resources. The code and dataset is available at https://github.com/Karl1109/CrackSCF.

An End-to-End Model for Time Series Classification In the Presence of Missing Values

Time series classification with missing data is a prevalent issue in time series analysis, as temporal data often contain missing values in practical applications. The traditional two-stage approach, which handles imputation and classification separately, can result in sub-optimal performance as label information is not utilized in the imputation process. On the other hand, a one-stage approach can learn features under missing information, but feature representation is limited as imputed errors are propagated in the classification process. To overcome these challenges, this study proposes an end-to-end neural network that unifies data imputation and representation learning within a single framework, allowing the imputation process to take advantage of label information. Differing from previous methods, our approach places less emphasis on the accuracy of imputation data and instead prioritizes classification performance. A specifically designed multi-scale feature learning module is implemented to extract useful information from the noise-imputation data. The proposed model is evaluated on 68 univariate time series datasets from the UCR archive, as well as a multivariate time series dataset with various missing data ratios and 4 real-world datasets with missing information. The results indicate that the proposed model outperforms state-of-the-art approaches for incomplete time series classification, particularly in scenarios with high levels of missing data.

Generalizing Graph Transformers Across Diverse Graphs and Tasks via Pre-Training on Industrial-Scale Data

Graph pre-training has been concentrated on graph-level on small graphs (e.g., molecular graphs) or learning node representations on a fixed graph. Extending graph pre-trained models to web-scale graphs with billions of nodes in industrial scenarios, while avoiding negative transfer across graphs or tasks, remains a challenge. We aim to develop a general graph pre-trained model with inductive ability that can make predictions for unseen new nodes and even new graphs. In this work, we introduce a scalable transformer-based graph pre-training framework called PGT (Pre-trained Graph Transformer). Specifically, we design a flexible and scalable graph transformer as the backbone network. Meanwhile, based on the masked autoencoder architecture, we design two pre-training tasks: one for reconstructing node features and the other one for reconstructing local structures. Unlike the original autoencoder architecture where the pre-trained decoder is discarded, we propose a novel strategy that utilizes the decoder for feature augmentation. We have deployed our framework on Tencent's online game data. Extensive experiments have demonstrated that our framework can perform pre-training on real-world web-scale graphs with over 540 million nodes and 12 billion edges and generalizes effectively to unseen new graphs with different downstream tasks. We further conduct experiments on the publicly available ogbn-papers100M dataset, which consists of 111 million nodes and 1.6 billion edges. Our framework achieves state-of-the-art performance on both industrial datasets and public datasets, while also enjoying scalability and efficiency.

Soft Label PU Learning

PU learning refers to the classification problem in which only part of positive samples are labeled. Existing PU learning methods treat unlabeled samples equally. However, in many real tasks, from common sense or domain knowledge, some unlabeled samples are more likely to be positive than others. In this paper, we propose soft label PU learning, in which unlabeled data are assigned soft labels according to their probabilities of being positive. Considering that the ground truth of TPR, FPR, and AUC are unknown, we then design PU counterparts of these metrics to evaluate the performances of soft label PU learning methods within validation data. We show that these new designed PU metrics are good substitutes for the real metrics. After that, a method that optimizes such metrics is proposed. Experiments on public datasets and real datasets for anti-cheat services from Tencent games demonstrate the effectiveness of our proposed method.

AdaFSNet: Time Series Classification Based on Convolutional Network with a Adaptive and Effective Kernel Size Configuration

Time series classification is one of the most critical and challenging problems in data mining, existing widely in various fields and holding significant research importance. Despite extensive research and notable achievements with successful real-world applications, addressing the challenge of capturing the appropriate receptive field (RF) size from one-dimensional or multi-dimensional time series of varying lengths remains a persistent issue, which greatly impacts performance and varies considerably across different datasets. In this paper, we propose an Adaptive and Effective Full-Scope Convolutional Neural Network (AdaFSNet) to enhance the accuracy of time series classification. This network includes two Dense Blocks. Particularly, it can dynamically choose a range of kernel sizes that effectively encompass the optimal RF size for various datasets by incorporating multiple prime numbers corresponding to the time series length. We also design a TargetDrop block, which can reduce redundancy while extracting a more effective RF. To assess the effectiveness of the AdaFSNet network, comprehensive experiments were conducted using the UCR and UEA datasets, which include one-dimensional and multi-dimensional time series data, respectively. Our model surpassed baseline models in terms of classification accuracy, underscoring the AdaFSNet network's efficiency and effectiveness in handling time series classification tasks.

CVPR 2023 Text Guided Video Editing Competition

Humans watch more than a billion hours of video per day. Most of this video was edited manually, which is a tedious process. However, AI-enabled video-generation and video-editing is on the rise. Building on text-to-image models like Stable Diffusion and Imagen, generative AI has improved dramatically on video tasks. But it's hard to evaluate progress in these video tasks because there is no standard benchmark. So, we propose a new dataset for text-guided video editing (TGVE), and we run a competition at CVPR to evaluate models on our TGVE dataset. In this paper we present a retrospective on the competition and describe the winning method. The competition dataset is available at https://sites.google.com/view/loveucvpr23/track4.

Modality Unifying Network for Visible-Infrared Person Re-Identification

Visible-infrared person re-identification (VI-ReID) is a challenging task due to large cross-modality discrepancies and intra-class variations. Existing methods mainly focus on learning modality-shared representations by embedding different modalities into the same feature space. As a result, the learned feature emphasizes the common patterns across modalities while suppressing modality-specific and identity-aware information that is valuable for Re-ID. To address these issues, we propose a novel Modality Unifying Network (MUN) to explore a robust auxiliary modality for VI-ReID. First, the auxiliary modality is generated by combining the proposed cross-modality learner and intra-modality learner, which can dynamically model the modality-specific and modality-shared representations to alleviate both cross-modality and intra-modality variations. Second, by aligning identity centres across the three modalities, an identity alignment loss function is proposed to discover the discriminative feature representations. Third, a modality alignment loss is introduced to consistently reduce the distribution distance of visible and infrared images by modality prototype modeling. Extensive experiments on multiple public datasets demonstrate that the proposed method surpasses the current state-of-the-art methods by a significant margin.