Most existing federated learning algorithms are based on the vanilla FedAvg scheme. However, with the increase of data complexity and the number of model parameters, the amount of communication traffic and the number of iteration rounds for training such algorithms increases significantly, especially in non-independently and homogeneously distributed scenarios, where they do not achieve satisfactory performance. In this work, we propose FedND: federated learning with noise distillation. The main idea is to use knowledge distillation to optimize the model training process. In the client, we propose a self-distillation method to train the local model. In the server, we generate noisy samples for each client and use them to distill other clients. Finally, the global model is obtained by the aggregation of local models. Experimental results show that the algorithm achieves the best performance and is more communication-efficient than state-of-the-art methods.
Machine learning-based forecasting models are commonly used in Intelligent Transportation Systems (ITS) to predict traffic patterns and provide city-wide services. However, most of the existing models are susceptible to adversarial attacks, which can lead to inaccurate predictions and negative consequences such as congestion and delays. Therefore, improving the adversarial robustness of these models is crucial for ITS. In this paper, we propose a novel framework for incorporating adversarial training into spatiotemporal traffic forecasting tasks. We demonstrate that traditional adversarial training methods designated for static domains cannot be directly applied to traffic forecasting tasks, as they fail to effectively defend against dynamic adversarial attacks. Then, we propose a reinforcement learning-based method to learn the optimal node selection strategy for adversarial examples, which simultaneously strengthens the dynamic attack defense capability and reduces the model overfitting. Additionally, we introduce a self-knowledge distillation regularization module to overcome the "forgetting issue" caused by continuously changing adversarial nodes during training. We evaluate our approach on two real-world traffic datasets and demonstrate its superiority over other baselines. Our method effectively enhances the adversarial robustness of spatiotemporal traffic forecasting models. The source code for our framework is available at https://github.com/usail-hkust/RDAT.
Scene information plays a crucial role in trajectory forecasting systems for autonomous driving by providing semantic clues and constraints on potential future paths of traffic agents. Prevalent trajectory prediction techniques often take high-definition maps (HD maps) as part of the inputs to provide scene knowledge. Although HD maps offer accurate road information, they may suffer from the high cost of annotation or restrictions of law that limits their widespread use. Therefore, those methods are still expected to generate reliable prediction results in mapless scenarios. In this paper, we tackle the problem of improving the consistency of multi-modal prediction trajectories and the real road topology when map information is unavailable during the test phase. Specifically, we achieve this by training a map-based prediction teacher network on the annotated samples and transferring the knowledge to a student mapless prediction network using a two-fold knowledge distillation framework. Our solution is generalizable for common trajectory prediction networks and does not bring extra computation burden. Experimental results show that our method stably improves prediction performance in mapless mode on many widely used state-of-the-art trajectory prediction baselines, compensating for the gaps caused by the absence of HD maps. Qualitative visualization results demonstrate that our approach helps infer unseen map information.
Transformers have emerged as the superior choice for face recognition tasks, but their insufficient platform acceleration hinders their application on mobile devices. In contrast, Convolutional Neural Networks (CNNs) capitalize on hardware-compatible acceleration libraries. Consequently, it has become indispensable to preserve the distillation efficacy when transferring knowledge from a Transformer-based teacher model to a CNN-based student model, known as Cross-Architecture Knowledge Distillation (CAKD). Despite its potential, the deployment of CAKD in face recognition encounters two challenges: 1) the teacher and student share disparate spatial information for each pixel, obstructing the alignment of feature space, and 2) the teacher network is not trained in the role of a teacher, lacking proficiency in handling distillation-specific knowledge. To surmount these two constraints, 1) we first introduce a Unified Receptive Fields Mapping module (URFM) that maps pixel features of the teacher and student into local features with unified receptive fields, thereby synchronizing the pixel-wise spatial information of teacher and student. Subsequently, 2) we develop an Adaptable Prompting Teacher network (APT) that integrates prompts into the teacher, enabling it to manage distillation-specific knowledge while preserving the model's discriminative capacity. Extensive experiments on popular face benchmarks and two large-scale verification sets demonstrate the superiority of our method.
Due to the point cloud's irregular and unordered geometry structure, conventional knowledge distillation technology lost a lot of information when directly used on point cloud tasks. In this paper, we propose Feature Adversarial Distillation (FAD) method, a generic adversarial loss function in point cloud distillation, to reduce loss during knowledge transfer. In the feature extraction stage, the features extracted by the teacher are used as the discriminator, and the students continuously generate new features in the training stage. The feature of the student is obtained by attacking the feedback from the teacher and getting a score to judge whether the student has learned the knowledge well or not. In experiments on standard point cloud classification on ModelNet40 and ScanObjectNN datasets, our method reduced the information loss of knowledge transfer in distillation in 40x model compression while maintaining competitive performance.
Knowledge distillation (KD) is widely used for compressing a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, current KD methods for auto-regressive sequence models suffer from distribution mismatch between output sequences seen during training and those generated by the student during inference. To address this issue, we introduce Generalized Knowledge Distillation (GKD). Instead of solely relying on a fixed set of output sequences, GKD trains the student on its self-generated output sequences by leveraging feedback from the teacher on such sequences. Unlike supervised KD approaches, GKD also offers the flexibility to employ alternative loss functions between the student and teacher, which can be useful when the student lacks the expressivity to mimic the teacher's distribution. Furthermore, GKD facilitates the seamless integration of distillation with RL fine-tuning (RLHF). We demonstrate the efficacy of GKD for distilling auto-regressive language models on summarization, translation, and arithmetic reasoning tasks, and task-agnostic distillation for instruction-tuning.
Abstract Meaning Representation (AMR) is a Semantic Parsing formalism that aims at providing a semantic graph abstraction representing a given text. Current approaches are based on autoregressive language models such as BART or T5, fine-tuned through Teacher Forcing to obtain a linearized version of the AMR graph from a sentence. In this paper, we present LeakDistill, a model and method that explores a modification to the Transformer architecture, using structural adapters to explicitly incorporate graph information into the learned representations and improve AMR parsing performance. Our experiments show how, by employing word-to-node alignment to embed graph structural information into the encoder at training time, we can obtain state-of-the-art AMR parsing through self-knowledge distillation, even without the use of additional data. We release the code at \url{http://www.github.com/sapienzanlp/LeakDistill}.
Knowledge distillation is a powerful technique for transferring knowledge from a pre-trained teacher model to a student model. However, the true potential of knowledge transfer has not been fully explored. Existing approaches primarily focus on distilling individual information or instance-level relationships, overlooking the valuable information embedded in token-level relationships, which may be particularly affected by the long-tail effects. To address the above limitations, we propose a novel method called Knowledge Distillation with Token-level Relationship Graph (TRG) that leverages the token-wise relational knowledge to enhance the performance of knowledge distillation. By employing TRG, the student model can effectively emulate higher-level semantic information from the teacher model, resulting in improved distillation results. To further enhance the learning process, we introduce a token-wise contextual loss called contextual loss, which encourages the student model to capture the inner-instance semantic contextual of the teacher model. We conduct experiments to evaluate the effectiveness of the proposed method against several state-of-the-art approaches. Empirical results demonstrate the superiority of TRG across various visual classification tasks, including those involving imbalanced data. Our method consistently outperforms the existing baselines, establishing a new state-of-the-art performance in the field of knowledge distillation.
Recently, speech-to-text translation has attracted more and more attention and many studies have emerged rapidly. In this paper, we present a comprehensive survey on direct speech translation aiming to summarize the current state-of-the-art techniques. First, we categorize the existing research work into three directions based on the main challenges -- modeling burden, data scarcity, and application issues. To tackle the problem of modeling burden, two main structures have been proposed, encoder-decoder framework (Transformer and the variants) and multitask frameworks. For the challenge of data scarcity, recent work resorts to many sophisticated techniques, such as data augmentation, pre-training, knowledge distillation, and multilingual modeling. We analyze and summarize the application issues, which include real-time, segmentation, named entity, gender bias, and code-switching. Finally, we discuss some promising directions for future work.
Existing skeleton-based action recognition methods typically follow a centralized learning paradigm, which can pose privacy concerns when exposing human-related videos. Federated Learning (FL) has attracted much attention due to its outstanding advantages in privacy-preserving. However, directly applying FL approaches to skeleton videos suffers from unstable training. In this paper, we investigate and discover that the heterogeneous human topology graph structure is the crucial factor hindering training stability. To address this limitation, we pioneer a novel Federated Skeleton-based Action Recognition (FSAR) paradigm, which enables the construction of a globally generalized model without accessing local sensitive data. Specifically, we introduce an Adaptive Topology Structure (ATS), separating generalization and personalization by learning a domain-invariant topology shared across clients and a domain-specific topology decoupled from global model aggregation.Furthermore, we explore Multi-grain Knowledge Distillation (MKD) to mitigate the discrepancy between clients and server caused by distinct updating patterns through aligning shallow block-wise motion features. Extensive experiments on multiple datasets demonstrate that FSAR outperforms state-of-the-art FL-based methods while inherently protecting privacy.
Deep neural networks have achieved remarkable performance for artificial intelligence tasks. The success behind intelligent systems often relies on large-scale models with high computational complexity and storage costs. The over-parameterized networks are often easy to optimize and can achieve better performance. However, it is challenging to deploy them over resource-limited edge-devices. Knowledge Distillation (KD) aims to optimize a lightweight network from the perspective of over-parameterized training. The traditional offline KD transfers knowledge from a cumbersome teacher to a small and fast student network. When a sizeable pre-trained teacher network is unavailable, online KD can improve a group of models by collaborative or mutual learning. Without needing extra models, Self-KD boosts the network itself using attached auxiliary architectures. KD mainly involves knowledge extraction and distillation strategies these two aspects. Beyond KD schemes, various KD algorithms are widely used in practical applications, such as multi-teacher KD, cross-modal KD, attention-based KD, data-free KD and adversarial KD. This paper provides a comprehensive KD survey, including knowledge categories, distillation schemes and algorithms, as well as some empirical studies on performance comparison. Finally, we discuss the open challenges of existing KD works and prospect the future directions.
Knowledge Distillation (KD) has been validated as an effective model compression technique for learning compact object detectors. Existing state-of-the-art KD methods for object detection are mostly based on feature imitation. In this paper, we present a general and effective prediction mimicking distillation scheme, called CrossKD, which delivers the intermediate features of the student's detection head to the teacher's detection head. The resulting cross-head predictions are then forced to mimic the teacher's predictions. This manner relieves the student's head from receiving contradictory supervision signals from the annotations and the teacher's predictions, greatly improving the student's detection performance. Moreover, as mimicking the teacher's predictions is the target of KD, CrossKD offers more task-oriented information in contrast with feature imitation. On MS COCO, with only prediction mimicking losses applied, our CrossKD boosts the average precision of GFL ResNet-50 with 1x training schedule from 40.2 to 43.7, outperforming all existing KD methods. In addition, our method also works well when distilling detectors with heterogeneous backbones. Code is available at https://github.com/jbwang1997/CrossKD.
Electrocardiography (ECG) is a non-invasive tool for predicting cardiovascular diseases (CVDs). Current ECG-based diagnosis systems show promising performance owing to the rapid development of deep learning techniques. However, the label scarcity problem, the co-occurrence of multiple CVDs and the poor performance on unseen datasets greatly hinder the widespread application of deep learning-based models. Addressing them in a unified framework remains a significant challenge. To this end, we propose a multi-label semi-supervised model (ECGMatch) to recognize multiple CVDs simultaneously with limited supervision. In the ECGMatch, an ECGAugment module is developed for weak and strong ECG data augmentation, which generates diverse samples for model training. Subsequently, a hyperparameter-efficient framework with neighbor agreement modeling and knowledge distillation is designed for pseudo-label generation and refinement, which mitigates the label scarcity problem. Finally, a label correlation alignment module is proposed to capture the co-occurrence information of different CVDs within labeled samples and propagate this information to unlabeled samples. Extensive experiments on four datasets and three protocols demonstrate the effectiveness and stability of the proposed model, especially on unseen datasets. As such, this model can pave the way for diagnostic systems that achieve robust performance on multi-label CVDs prediction with limited supervision.
Automatically describing audio-visual content with texts, namely video captioning, has received significant attention due to its potential applications across diverse fields. Deep neural networks are the dominant methods, offering state-of-the-art performance. However, these methods are often undeployable in low-power devices like smartphones due to the large size of the model parameters. In this paper, we propose to exploit simple pooling front-end and down-sampling algorithms with knowledge distillation for audio and visual attributes using a reduced number of audio-visual frames. With the help of knowledge distillation from the teacher model, our proposed method greatly reduces the redundant information in audio-visual streams without losing critical contexts for caption generation. Extensive experimental evaluations on the MSR-VTT dataset demonstrate that our proposed approach significantly reduces the inference time by about 80% with a small sacrifice (less than 0.02%) in captioning accuracy.
Cloud detection is a pivotal satellite image pre-processing step that can be performed both on the ground and on board a satellite to tag useful images. In the latter case, it can reduce the amount of data to downlink by pruning the cloudy areas, or to make a satellite more autonomous through data-driven acquisition re-scheduling. We approach this task with nnU-Nets, a self-reconfigurable framework able to perform meta-learning of a segmentation network over various datasets. Unfortunately, such models are commonly memory-inefficient due to their (very) large architectures. To benefit from them in on-board processing, we compress nnU-Nets with knowledge distillation into much smaller and compact U-Nets. Our experiments, performed over Sentinel-2 and Landsat-8 images revealed that nnU-Nets deliver state-of-the-art performance without any manual design. Our approach was ranked within the top 7% best solutions (across 847 teams) in the On Cloud N: Cloud Cover Detection Challenge, where we reached the Jaccard index of 0.882 over more than 10k unseen Sentinel-2 images (the winners obtained 0.897, the baseline U-Net with the ResNet-34 backbone: 0.817, and the classic Sentinel-2 image thresholding: 0.652). Finally, we showed that knowledge distillation enables to elaborate dramatically smaller (almost 280x) U-Nets when compared to nnU-Nets while still maintaining their segmentation capabilities.
For a very long time, unsupervised learning for anomaly detection has been at the heart of image processing research and a stepping stone for high performance industrial automation process. With the emergence of CNN, several methods have been proposed such as Autoencoders, GAN, deep feature extraction, etc. In this paper, we propose a new method based on the promising concept of knowledge distillation which consists of training a network (the student) on normal samples while considering the output of a larger pretrained network (the teacher). The main contributions of this paper are twofold: First, a reduced student architecture with optimal layer selection is proposed, then a new Student-Teacher architecture with network bias reduction combining two teachers is proposed in order to jointly enhance the performance of anomaly detection and its localization accuracy. The proposed texture anomaly detector has an outstanding capability to detect defects in any texture and a fast inference time compared to the SOTA methods.
We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Current distillation for sensorimotor agents methods tend to result in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key insight is to design a student which learns to align their input features with the teacher's privileged Bird's Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mechanism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that surpasses prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sensorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over 20.6% in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning.
Current knowledge distillation approaches in semantic segmentation tend to adopt a holistic approach that treats all spatial locations equally. However, for dense prediction, students' predictions on edge regions are highly uncertain due to contextual information leakage, requiring higher spatial sensitivity knowledge than the body regions. To address this challenge, this paper proposes a novel approach called boundary-privileged knowledge distillation (BPKD). BPKD distills the knowledge of the teacher model's body and edges separately to the compact student model. Specifically, we employ two distinct loss functions: (i) edge loss, which aims to distinguish between ambiguous classes at the pixel level in edge regions; (ii) body loss, which utilizes shape constraints and selectively attends to the inner-semantic regions. Our experiments demonstrate that the proposed BPKD method provides extensive refinements and aggregation for edge and body regions. Additionally, the method achieves state-of-the-art distillation performance for semantic segmentation on three popular benchmark datasets, highlighting its effectiveness and generalization ability. BPKD shows consistent improvements across a diverse array of lightweight segmentation structures, including both CNNs and transformers, underscoring its architecture-agnostic adaptability. The code is available at \url{https://github.com/AkideLiu/BPKD}.
Conventional knowledge distillation (KD) methods require access to the internal information of teachers, e.g., logits. However, such information may not always be accessible for large pre-trained language models (PLMs). In this work, we focus on decision-based KD for PLMs, where only teacher decisions (i.e., top-1 labels) are accessible. Considering the information gap between logits and decisions, we propose a novel method to estimate logits from the decision distributions. Specifically, decision distributions can be both derived as a function of logits theoretically and estimated with test-time data augmentation empirically. By combining the theoretical and empirical estimations of the decision distributions together, the estimation of logits can be successfully reduced to a simple root-finding problem. Extensive experiments show that our method significantly outperforms strong baselines on both natural language understanding and machine reading comprehension datasets.
Knowledge Distillation (KD) is a promising technique for reducing the high computational demand of large language models (LLMs). However, previous KD methods are primarily applied to white-box classification models or training small models to imitate black-box model APIs like ChatGPT. How to effectively distill the knowledge of white-box LLMs into small models is still under-explored, which becomes more important with the prosperity of open-source LLMs. In this work, we propose a KD approach that distills LLMs into smaller language models. We first replace the forward Kullback-Leibler divergence (KLD) objective in the standard KD approaches with reverse KLD, which is more suitable for KD on generative language models, to prevent the student model from overestimating the low-probability regions of the teacher distribution. Then, we derive an effective optimization approach to learn this objective. The student models are named MiniLLM. Extensive experiments in the instruction-following setting show that MiniLLM generates more precise responses with higher overall quality, lower exposure bias, better calibration, and higher long-text generation performance than the baselines. Our method is scalable for different model families with 120M to 13B parameters. Our code, data, and model checkpoints can be found in https://github.com/microsoft/LMOps/tree/main/minillm.