Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD.
Existing knowledge distillation methods generally use a teacher-student approach, where the student network solely learns from a well-trained teacher. However, this approach overlooks the inherent differences in learning abilities between the teacher and student networks, thus causing the capacity-gap problem. To address this limitation, we propose a novel method called SLKD.
Cooperative perception can significantly improve the perception performance of autonomous vehicles beyond the limited perception ability of individual vehicles by exchanging information with neighbor agents through V2X communication. However, most existing work assume ideal communication among agents, ignoring the significant and common \textit{interruption issues} caused by imperfect V2X communication, where cooperation agents can not receive cooperative messages successfully and thus fail to achieve cooperative perception, leading to safety risks. To fully reap the benefits of cooperative perception in practice, we propose V2X communication INterruption-aware COoperative Perception (V2X-INCOP), a cooperative perception system robust to communication interruption for V2X communication-aided autonomous driving, which leverages historical cooperation information to recover missing information due to the interruptions and alleviate the impact of the interruption issue. To achieve comprehensive recovery, we design a communication-adaptive multi-scale spatial-temporal prediction model to extract multi-scale spatial-temporal features based on V2X communication conditions and capture the most significant information for the prediction of the missing information. To further improve recovery performance, we adopt a knowledge distillation framework to give explicit and direct supervision to the prediction model and a curriculum learning strategy to stabilize the training of the model. Experiments on three public cooperative perception datasets demonstrate that the proposed method is effective in alleviating the impacts of communication interruption on cooperative perception.
LiDAR point cloud segmentation is one of the most fundamental tasks for autonomous driving scene understanding. However, it is difficult for existing models to achieve both high inference speed and accuracy simultaneously. For example, voxel-based methods perform well in accuracy, while Bird's-Eye-View (BEV)-based methods can achieve real-time inference. To overcome this issue, we develop an effective 3D-to-BEV knowledge distillation method that transfers rich knowledge from 3D voxel-based models to BEV-based models. Our framework mainly consists of two modules: the voxel-to-pillar distillation module and the label-weight distillation module. Voxel-to-pillar distillation distills sparse 3D features to BEV features for middle layers to make the BEV-based model aware of more structural and geometric information. Label-weight distillation helps the model pay more attention to regions with more height information. Finally, we conduct experiments on the SemanticKITTI dataset and Paris-Lille-3D. The results on SemanticKITTI show more than 5% improvement on the test set, especially for classes such as motorcycle and person, with more than 15% improvement. The code can be accessed at https://github.com/fengjiang5/Knowledge-Distillation-from-Cylinder3D-to-PolarNet.
Automatic Pill Recognition (APR) systems are crucial for enhancing hospital efficiency, assisting visually impaired individuals, and preventing cross-infection. However, most existing deep learning-based pill recognition systems can only perform classification on classes with sufficient training data. In practice, the high cost of data annotation and the continuous increase in new pill classes necessitate the development of a few-shot class-incremental pill recognition system. This paper introduces the first few-shot class-incremental pill recognition framework, named Discriminative and Bidirectional Compatible Few-Shot Class-Incremental Learning (DBC-FSCIL). It encompasses forward-compatible and backward-compatible learning components. In forward-compatible learning, we propose an innovative virtual class synthesis strategy and a Center-Triplet (CT) loss to enhance discriminative feature learning. These virtual classes serve as placeholders in the feature space for future class updates, providing diverse semantic knowledge for model training. For backward-compatible learning, we develop a strategy to synthesize reliable pseudo-features of old classes using uncertainty quantification, facilitating Data Replay (DR) and Knowledge Distillation (KD). This approach allows for the flexible synthesis of features and effectively reduces additional storage requirements for samples and models. Additionally, we construct a new pill image dataset for FSCIL and assess various mainstream FSCIL methods, establishing new benchmarks. Our experimental results demonstrate that our framework surpasses existing State-of-the-art (SOTA) methods. The code is available at https://github.com/zhang-jinghua/DBC-FSCIL.
Knowledge distillation constitutes a potent methodology for condensing substantial neural networks into more compact and efficient counterparts. Within this context, softmax regression representation learning serves as a widely embraced approach, leveraging a pre-established teacher network to guide the learning process of a diminutive student network. Notably, despite the extensive inquiry into the efficacy of softmax regression representation learning, the intricate underpinnings governing the knowledge transfer mechanism remain inadequately elucidated. This study introduces the 'Ideal Joint Classifier Knowledge Distillation' (IJCKD) framework, an overarching paradigm that not only furnishes a lucid and exhaustive comprehension of prevailing knowledge distillation techniques but also establishes a theoretical underpinning for prospective investigations. Employing mathematical methodologies derived from domain adaptation theory, this investigation conducts a comprehensive examination of the error boundary of the student network contingent upon the teacher network. Consequently, our framework facilitates efficient knowledge transference between teacher and student networks, thereby accommodating a diverse spectrum of applications.
How can we learn effective node representations on textual graphs? Graph Neural Networks (GNNs) that use Language Models (LMs) to encode textual information of graphs achieve state-of-the-art performance in many node classification tasks. Yet, combining GNNs with LMs has not been widely explored for practical deployments due to its scalability issues. In this work, we tackle this challenge by developing a Graph-Aware Distillation framework (GRAD) to encode graph structures into an LM for graph-free, fast inference. Different from conventional knowledge distillation, GRAD jointly optimizes a GNN teacher and a graph-free student over the graph's nodes via a shared LM. This encourages the graph-free student to exploit graph information encoded by the GNN teacher while at the same time, enables the GNN teacher to better leverage textual information from unlabeled nodes. As a result, the teacher and the student models learn from each other to improve their overall performance. Experiments in eight node classification benchmarks in both transductive and inductive settings showcase GRAD's superiority over existing distillation approaches for textual graphs.
Pre-trained contextual language models are ubiquitously employed for language understanding tasks, but are unsuitable for resource-constrained systems. Noncontextual word embeddings are an efficient alternative in these settings. Such methods typically use one vector to encode multiple different meanings of a word, and incur errors due to polysemy. This paper proposes a two-stage method to distill multiple word senses from a pre-trained language model (BERT) by using attention over the senses of a word in a context and transferring this sense information to fit multi-sense embeddings in a skip-gram-like framework. We demonstrate an effective approach to training the sense disambiguation mechanism in our model with a distribution over word senses extracted from the output layer embeddings of BERT. Experiments on the contextual word similarity and sense induction tasks show that this method is superior to or competitive with state-of-the-art multi-sense embeddings on multiple benchmark data sets, and experiments with an embedding-based topic model (ETM) demonstrates the benefits of using this multi-sense embedding in a downstream application.
Existing techniques often attempt to make knowledge transfer from a powerful machine translation (MT) to speech translation (ST) model with some elaborate techniques, which often requires transcription as extra input during training. However, transcriptions are not always available, and how to improve the ST model performance without transcription, i.e., data efficiency, has rarely been studied in the literature. In this paper, we propose Decoupled Non-parametric Knowledge Distillation (DNKD) from data perspective to improve the data efficiency. Our method follows the knowledge distillation paradigm. However, instead of obtaining the teacher distribution from a sophisticated MT model, we construct it from a non-parametric datastore via k-Nearest-Neighbor (kNN) retrieval, which removes the dependence on transcription and MT model. Then we decouple the classic knowledge distillation loss into target and non-target distillation to enhance the effect of the knowledge among non-target logits, which is the prominent "dark knowledge". Experiments on MuST-C corpus show that, the proposed method can achieve consistent improvement over the strong baseline without requiring any transcription.
Spiking neural networks (SNNs) have superb characteristics in sensory information recognition tasks due to their biological plausibility. However, the performance of some current spiking-based models is limited by their structures which means either fully connected or too-deep structures bring too much redundancy. This redundancy from both connection and neurons is one of the key factors hindering the practical application of SNNs. Although Some pruning methods were proposed to tackle this problem, they normally ignored the fact the neural topology in the human brain could be adjusted dynamically. Inspired by this, this paper proposed an evolutionary-based structure construction method for constructing more reasonable SNNs. By integrating the knowledge distillation and connection pruning method, the synaptic connections in SNNs can be optimized dynamically to reach an optimal state. As a result, the structure of SNNs could not only absorb knowledge from the teacher model but also search for deep but sparse network topology. Experimental results on CIFAR100 and DVS-Gesture show that the proposed structure learning method can get pretty well performance while reducing the connection redundancy. The proposed method explores a novel dynamical way for structure learning from scratch in SNNs which could build a bridge to close the gap between deep learning and bio-inspired neural dynamics.
Knowledge distillation (KD) is a well-known method for compressing neural models. However, works focusing on distilling knowledge from large multilingual neural machine translation (MNMT) models into smaller ones are practically nonexistent, despite the popularity and superiority of MNMT. This paper bridges this gap by presenting an empirical investigation of knowledge distillation for compressing MNMT models. We take Indic to English translation as a case study and demonstrate that commonly used language-agnostic and language-aware KD approaches yield models that are 4-5x smaller but also suffer from performance drops of up to 3.5 BLEU. To mitigate this, we then experiment with design considerations such as shallower versus deeper models, heavy parameter sharing, multi-stage training, and adapters. We observe that deeper compact models tend to be as good as shallower non-compact ones, and that fine-tuning a distilled model on a High-Quality subset slightly boosts translation quality. Overall, we conclude that compressing MNMT models via KD is challenging, indicating immense scope for further research.
Spiking Neural Networks (SNNs) are biologically realistic and practically promising in low-power computation because of their event-driven mechanism. Usually, the training of SNNs suffers accuracy loss on various tasks, yielding an inferior performance compared with ANNs. A conversion scheme is proposed to obtain competitive accuracy by mapping trained ANNs' parameters to SNNs with the same structures. However, an enormous number of time steps are required for these converted SNNs, thus losing the energy-efficient benefit. Utilizing both the accuracy advantages of ANNs and the computing efficiency of SNNs, a novel SNN training framework is proposed, namely layer-wise ANN-to-SNN knowledge distillation (LaSNN). In order to achieve competitive accuracy and reduced inference latency, LaSNN transfers the learning from a well-trained ANN to a small SNN by distilling the knowledge other than converting the parameters of ANN. The information gap between heterogeneous ANN and SNN is bridged by introducing the attention scheme, the knowledge in an ANN is effectively compressed and then efficiently transferred by utilizing our layer-wise distillation paradigm. We conduct detailed experiments to demonstrate the effectiveness, efficacy, and scalability of LaSNN on three benchmark data sets (CIFAR-10, CIFAR-100, and Tiny ImageNet). We achieve competitive top-1 accuracy compared to ANNs and 20x faster inference than converted SNNs with similar performance. More importantly, LaSNN is dexterous and extensible that can be effortlessly developed for SNNs with different architectures/depths and input encoding methods, contributing to their potential development.
Click-through rate (CTR) prediction is of great importance in recommendation systems and online advertising platforms. When served in industrial scenarios, the user-generated data observed by the CTR model typically arrives as a stream. Streaming data has the characteristic that the underlying distribution drifts over time and may recur. This can lead to catastrophic forgetting if the model simply adapts to new data distribution all the time. Also, it's inefficient to relearn distribution that has been occurred. Due to memory constraints and diversity of data distributions in large-scale industrial applications, conventional strategies for catastrophic forgetting such as replay, parameter isolation, and knowledge distillation are difficult to be deployed. In this work, we design a novel drift-aware incremental learning framework based on ensemble learning to address catastrophic forgetting in CTR prediction. With explicit error-based drift detection on streaming data, the framework further strengthens well-adapted ensembles and freezes ensembles that do not match the input distribution avoiding catastrophic interference. Both evaluations on offline experiments and A/B test shows that our method outperforms all baselines considered.
Many existing studies on knowledge distillation have focused on methods in which a student model mimics a teacher model well. Simply imitating the teacher's knowledge, however, is not sufficient for the student to surpass that of the teacher. We explore a method to harness the knowledge of other students to complement the knowledge of the teacher. We propose deep collective knowledge distillation for model compression, called DCKD, which is a method for training student models with rich information to acquire knowledge from not only their teacher model but also other student models. The knowledge collected from several student models consists of a wealth of information about the correlation between classes. Our DCKD considers how to increase the correlation knowledge of classes during training. Our novel method enables us to create better performing student models for collecting knowledge. This simple yet powerful method achieves state-of-the-art performances in many experiments. For example, for ImageNet, ResNet18 trained with DCKD achieves 72.27\%, which outperforms the pretrained ResNet18 by 2.52\%. For CIFAR-100, the student model of ShuffleNetV1 with DCKD achieves 6.55\% higher top-1 accuracy than the pretrained ShuffleNetV1.
The ability to recognize, localize and track dynamic objects in a scene is fundamental to many real-world applications, such as self-driving and robotic systems. Yet, traditional multiple object tracking (MOT) benchmarks rely only on a few object categories that hardly represent the multitude of possible objects that are encountered in the real world. This leaves contemporary MOT methods limited to a small set of pre-defined object categories. In this paper, we address this limitation by tackling a novel task, open-vocabulary MOT, that aims to evaluate tracking beyond pre-defined training categories. We further develop OVTrack, an open-vocabulary tracker that is capable of tracking arbitrary object classes. Its design is based on two key ingredients: First, leveraging vision-language models for both classification and association via knowledge distillation; second, a data hallucination strategy for robust appearance feature learning from denoising diffusion probabilistic models. The result is an extremely data-efficient open-vocabulary tracker that sets a new state-of-the-art on the large-scale, large-vocabulary TAO benchmark, while being trained solely on static images. Project page: https://www.vis.xyz/pub/ovtrack/
The Segment Anything Model (SAM) is a promptable segmentation model recently introduced by Meta AI that has demonstrated its prowess across various fields beyond just image segmentation. SAM can accurately segment images across diverse fields, and generating various masks. We discovered that this ability of SAM can be leveraged to pretrain models for specific fields. Accordingly, we have proposed a framework that utilizes SAM to generate pseudo labels for pretraining thermal infrared image segmentation tasks. Our proposed framework can effectively improve the accuracy of segmentation results of specific categories beyond the SOTA ImageNet pretrained model. Our framework presents a novel approach to collaborate with models trained with large data like SAM to address problems in special fields. Also, we generated a large scale thermal infrared segmentation dataset used for pretaining, which contains over 100,000 images with pixel-annotation labels. This approach offers an effective solution for working with large models in special fields where label annotation is challenging. Our code is available at https://github.com/chenjzBUAA/SATIR
Cross-modal distillation has been widely used to transfer knowledge across different modalities, enriching the representation of the target unimodal one. Recent studies highly relate the temporal synchronization between vision and sound to the semantic consistency for cross-modal distillation. However, such semantic consistency from the synchronization is hard to guarantee in unconstrained videos, due to the irrelevant modality noise and differentiated semantic correlation. To this end, we first propose a \textit{Modality Noise Filter} (MNF) module to erase the irrelevant noise in teacher modality with cross-modal context. After this purification, we then design a \textit{Contrastive Semantic Calibration} (CSC) module to adaptively distill useful knowledge for target modality, by referring to the differentiated sample-wise semantic correlation in a contrastive fashion. Extensive experiments show that our method could bring a performance boost compared with other distillation methods in both visual action recognition and video retrieval task. We also extend to the audio tagging task to prove the generalization of our method. The source code is available at \href{https://github.com/GeWu-Lab/cross-modal-distillation}{https://github.com/GeWu-Lab/cross-modal-distillation}.
We investigate cross-quality knowledge distillation (CQKD), a knowledge distillation method where knowledge from a teacher network trained with full-resolution images is transferred to a student network that takes as input low-resolution images. As image size is a deciding factor for the computational load of computer vision applications, CQKD notably reduces the requirements by only using the student network at inference time. Our experimental results show that CQKD outperforms supervised learning in large-scale image classification problems. We also highlight the importance of calibrating neural networks: we show that with higher temperature smoothing of the teacher's output distribution, the student distribution exhibits a higher entropy, which leads to both, a lower calibration error and a higher network accuracy.
Language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning has become the dominant paradigm in NLP. These pre-training datasets often have a one-to-many structure--e.g. in dialogue there are many valid responses for a given context. However, only some of these responses will be desirable in our downstream task. This raises the question of how we should train the model such that it can emulate the desirable behaviours, but not the undesirable ones. Current approaches train in a one-to-one setup--only a single target response is given for a single dialogue context--leading to models only learning to predict the average response, while ignoring the full range of possible responses. Using text-based games as a testbed, our approach, PASA, uses discrete latent variables to capture the range of different behaviours represented in our larger pre-training dataset. We then use knowledge distillation to distil the posterior probability distribution into a student model. This probability distribution is far richer than learning from only the hard targets of the dataset, and thus allows the student model to benefit from the richer range of actions the teacher model has learned. Results show up to 49% empirical improvement over the previous state-of-the-art model on the Jericho Walkthroughs dataset.
Wearable devices for seizure monitoring detection could significantly improve the quality of life of epileptic patients. However, existing solutions that mostly rely on full electrode set of electroencephalogram (EEG) measurements could be inconvenient for every day use. In this paper, we propose a novel knowledge distillation approach to transfer the knowledge from a sophisticated seizure detector (called the teacher) trained on data from the full set of electrodes to learn new detectors (called the student). They are both providing lightweight implementations and significantly reducing the number of electrodes needed for recording the EEG. We consider the case where the teacher and the student seizure detectors are graph neural networks (GNN), since these architectures actively use the connectivity information. We consider two cases (a) when a single student is learnt for all the patients using preselected channels; and (b) when personalized students are learnt for every individual patient, with personalized channel selection using a Gumbelsoftmax approach. Our experiments on the publicly available Temple University Hospital EEG Seizure Data Corpus (TUSZ) show that both knowledge-distillation and personalization play significant roles in improving performance of seizure detection, particularly for patients with scarce EEG data. We observe that using as few as two channels, we are able to obtain competitive seizure detection performance. This, in turn, shows the potential of our approach in more realistic scenario of wearable devices for personalized monitoring of seizures, even with few recordings.