A Low Earth orbit (LEO) satellite constellation consists of a large number of small satellites traveling in space with high mobility and collecting vast amounts of mobility data such as cloud movement for weather forecast, large herds of animals migrating across geo-regions, spreading of forest fires, and aircraft tracking. Machine learning can be utilized to analyze these mobility data to address global challenges, and Federated Learning (FL) is a promising approach because it eliminates the need for transmitting raw data and hence is both bandwidth and privacy-friendly. However, FL requires many communication rounds between clients (satellites) and the parameter server (PS), leading to substantial delays of up to several days in LEO constellations. In this paper, we propose a novel one-shot FL approach for LEO satellites, called LEOShot, that needs only a single communication round to complete the entire learning process. LEOShot comprises three processes: (i) synthetic data generation, (ii) knowledge distillation, and (iii) virtual model retraining. We evaluate and benchmark LEOShot against the state of the art and the results show that it drastically expedites FL convergence by more than an order of magnitude. Also surprisingly, despite the one-shot nature, its model accuracy is on par with or even outperforms regular iterative FL schemes by a large margin
Pretrained language models (LMs) have shown compelling performance on various downstream tasks, but unfortunately they require a tremendous amount of inference compute. Knowledge distillation finds a path to compress LMs to small ones with a teacher-student paradigm. However, when the capacity gap between the teacher and the student is large, a curse of capacity gap appears, invoking a deficiency in distilling LMs. While a few studies have been carried out to fill the gap, the curse is not yet well tackled. In this paper, we aim at lifting the curse of capacity gap via enlarging the capacity of the student without notably increasing the inference compute. Largely motivated by sparse activation regime of mixture of experts (MoE), we propose a mixture of minimal experts (MiniMoE), which imposes extra parameters to the student but introduces almost no additional inference compute. Experimental results on GLUE and CoNLL demonstrate the curse of capacity gap is lifted by the magic of MiniMoE to a large extent. MiniMoE also achieves the state-of-the-art performance at small FLOPs compared with a range of competitive baselines. With a compression rate as much as $\sim$50$\times$, MiniMoE preserves $\sim$95\% GLUE score of the teacher.
Knowledge distillation optimises a smaller student model to behave similarly to a larger teacher model, retaining some of the performance benefits. While this method can improve results on in-distribution examples, it does not necessarily generalise to out-of-distribution (OOD) settings. We investigate two complementary methods for improving the robustness of the resulting student models on OOD domains. The first approach augments the distillation with generated unlabelled examples that match the target distribution. The second method upsamples data points among the training set that are similar to the target distribution. When applied on the task of natural language inference (NLI), our experiments on MNLI show that distillation with these modifications outperforms previous robustness solutions. We also find that these methods improve performance on OOD domains even beyond the target domain.
The excellent performance of deep neural networks is usually accompanied by a large number of parameters and computations, which have limited their usage on the resource-limited edge devices. To address this issue, abundant methods such as pruning, quantization and knowledge distillation have been proposed to compress neural networks and achieved significant breakthroughs. However, most of these compression methods focus on the architecture or the training method of neural networks but ignore the influence from data augmentation. In this paper, we revisit the usage of data augmentation in model compression and give a comprehensive study on the relation between model sizes and their optimal data augmentation policy. To sum up, we mainly have the following three observations: (A) Models in different sizes prefer data augmentation with different magnitudes. Hence, in iterative pruning, data augmentation with varying magnitudes leads to better performance than data augmentation with a consistent magnitude. (B) Data augmentation with a high magnitude may significantly improve the performance of large models but harm the performance of small models. Fortunately, small models can still benefit from strong data augmentations by firstly learning them with "additional parameters" and then discard these "additional parameters" during inference. (C) The prediction of a pre-trained large model can be utilized to measure the difficulty of data augmentation. Thus it can be utilized as a criterion to design better data augmentation policies. We hope this paper may promote more research on the usage of data augmentation in model compression.
Diffusion models have recently achieved astonishing performance in generating high-fidelity photo-realistic images. Given their huge success, it is still unclear whether synthetic images are applicable for knowledge distillation when real images are unavailable. In this paper, we extensively study whether and how synthetic images produced from state-of-the-art diffusion models can be used for knowledge distillation without access to real images, and obtain three key conclusions: (1) synthetic data from diffusion models can easily lead to state-of-the-art performance among existing synthesis-based distillation methods, (2) low-fidelity synthetic images are better teaching materials, and (3) relatively weak classifiers are better teachers. Code is available at https://github.com/zhengli97/DM-KD.
Many-to-many multimodal summarization (M$^3$S) task aims to generate summaries in any language with document inputs in any language and the corresponding image sequence, which essentially comprises multimodal monolingual summarization (MMS) and multimodal cross-lingual summarization (MXLS) tasks. Although much work has been devoted to either MMS or MXLS and has obtained increasing attention in recent years, little research pays attention to the M$^3$S task. Besides, existing studies mainly focus on 1) utilizing MMS to enhance MXLS via knowledge distillation without considering the performance of MMS or 2) improving MMS models by filtering summary-unrelated visual features with implicit learning or explicitly complex training objectives. In this paper, we first introduce a general and practical task, i.e., M$^3$S. Further, we propose a dual knowledge distillation and target-oriented vision modeling framework for the M$^3$S task. Specifically, the dual knowledge distillation method guarantees that the knowledge of MMS and MXLS can be transferred to each other and thus mutually prompt both of them. To offer target-oriented visual features, a simple yet effective target-oriented contrastive objective is designed and responsible for discarding needless visual information. Extensive experiments on the many-to-many setting show the effectiveness of the proposed approach. Additionally, we will contribute a many-to-many multimodal summarization (M$^3$Sum) dataset.
The pre-trained speech encoder wav2vec 2.0 performs very well on various spoken language understanding (SLU) tasks. However, on many tasks, it trails behind text encoders with textual input. To improve the understanding capability of SLU encoders, various studies have used knowledge distillation to transfer knowledge from natural language understanding (NLU) encoders. We use a very simple method of distilling from a textual sentence embedder directly into wav2vec 2.0 as pre-training, utilizing paired audio-text datasets. We observed that this method is indeed capable of improving SLU task performance in fine-tuned settings, as well as full-data and few-shot transfer on a frozen encoder. However, the model performs worse on certain tasks highlighting the strengths and weaknesses of our approach.
Like many other machine learning applications, neural machine translation (NMT) benefits from over-parameterized deep neural models. However, these models have been observed to be brittle: NMT model predictions are sensitive to small input changes and can show significant variation across re-training or incremental model updates. This work studies a frequently used method in NMT, pseudo-label training (PLT), which is common to the related techniques of forward-translation (or self-training) and sequence-level knowledge distillation. While the effect of PLT on quality is well-documented, we highlight a lesser-known effect: PLT can enhance a model's stability to model updates and input perturbations, a set of properties we call model inertia. We study inertia effects under different training settings and we identify distribution simplification as a mechanism behind the observed results.
End-to-end sign language translation (SLT) aims to convert sign language videos into spoken language texts directly without intermediate representations. It has been a challenging task due to the modality gap between sign videos and texts and the data scarcity of labeled data. Due to these challenges, the input and output distributions of end-to-end sign language translation (i.e., video-to-text) are less effective compared to the gloss-to-text approach (i.e., text-to-text). To tackle these challenges, we propose a novel Cross-modality Data Augmentation (XmDA) framework to transfer the powerful gloss-to-text translation capabilities to end-to-end sign language translation (i.e. video-to-text) by exploiting pseudo gloss-text pairs from the sign gloss translation model. Specifically, XmDA consists of two key components, namely, cross-modality mix-up and cross-modality knowledge distillation. The former explicitly encourages the alignment between sign video features and gloss embeddings to bridge the modality gap. The latter utilizes the generation knowledge from gloss-to-text teacher models to guide the spoken language text generation. Experimental results on two widely used SLT datasets, i.e., PHOENIX-2014T and CSL-Daily, demonstrate that the proposed XmDA framework significantly and consistently outperforms the baseline models. Extensive analyses confirm our claim that XmDA enhances spoken language text generation by reducing the representation distance between videos and texts, as well as improving the processing of low-frequency words and long sentences.
Dense retrieval has shown promise in the first-stage retrieval process when trained on in-domain labeled datasets. However, previous studies have found that dense retrieval is hard to generalize to unseen domains due to its weak modeling of domain-invariant and interpretable feature (i.e., matching signal between two texts, which is the essence of information retrieval). In this paper, we propose a novel method to improve the generalization of dense retrieval via capturing matching signal called BERM. Fully fine-grained expression and query-oriented saliency are two properties of the matching signal. Thus, in BERM, a single passage is segmented into multiple units and two unit-level requirements are proposed for representation as the constraint in training to obtain the effective matching signal. One is semantic unit balance and the other is essential matching unit extractability. Unit-level view and balanced semantics make representation express the text in a fine-grained manner. Essential matching unit extractability makes passage representation sensitive to the given query to extract the pure matching information from the passage containing complex context. Experiments on BEIR show that our method can be effectively combined with different dense retrieval training methods (vanilla, hard negatives mining and knowledge distillation) to improve its generalization ability without any additional inference overhead and target domain data.
As a popular multilingual and multitask pre-trained speech model, Whisper has the problem of curse of multilinguality. To enhance multilingual capabilities in small Whisper models, we propose DQ-Whisper, a novel joint distillation and quantization framework to compress Whisper for efficient inference. Firstly, we propose a novel dynamic matching distillation strategy. Then, a quantization-aware distillation framework is introduced to integrate quantization with distillation. Experimental results on various multilingual datasets show that our suggested distillation approach can effectively enhance the multilingual capabilities of small Whisper models without increasing computational costs. Up to 5.18x reduction in model size is achieved with marginal performance degradation. In addition, quantization is compatible with distillation, which can result in a higher compression rate.
Diffusion Probability Models (DPMs) have made impressive advancements in various machine learning domains. However, achieving high-quality synthetic samples typically involves performing a large number of sampling steps, which impedes the possibility of real-time sample synthesis. Traditional accelerated sampling algorithms via knowledge distillation rely on pre-trained model weights and discrete time step scenarios, necessitating additional training sessions to achieve their goals. To address these issues, we propose the Catch-Up Distillation (CUD), which encourages the current moment output of the velocity estimation model ``catch up'' with its previous moment output. Specifically, CUD adjusts the original Ordinary Differential Equation (ODE) training objective to align the current moment output with both the ground truth label and the previous moment output, utilizing Runge-Kutta-based multi-step alignment distillation for precise ODE estimation while preventing asynchronous updates. Furthermore, we investigate the design space for CUDs under continuous time-step scenarios and analyze how to determine the suitable strategies. To demonstrate CUD's effectiveness, we conduct thorough ablation and comparison experiments on CIFAR-10, MNIST, and ImageNet-64. On CIFAR-10, we obtain a FID of 2.80 by sampling in 15 steps under one-session training and the new state-of-the-art FID of 3.37 by sampling in one step with additional training. This latter result necessitated only 620k iterations with a batch size of 128, in contrast to Consistency Distillation, which demanded 2100k iterations with a larger batch size of 256. Our code is released at https://anonymous.4open.science/r/Catch-Up-Distillation-E31F.
The transformer extends its success from the language to the vision domain. Because of the stacked self-attention and cross-attention blocks, the acceleration deployment of vision transformer on GPU hardware is challenging and also rarely studied. This paper thoroughly designs a compression scheme to maximally utilize the GPU-friendly 2:4 fine-grained structured sparsity and quantization. Specially, an original large model with dense weight parameters is first pruned into a sparse one by 2:4 structured pruning, which considers the GPU's acceleration of 2:4 structured sparse pattern with FP16 data type, then the floating-point sparse model is further quantized into a fixed-point one by sparse-distillation-aware quantization aware training, which considers GPU can provide an extra speedup of 2:4 sparse calculation with integer tensors. A mixed-strategy knowledge distillation is used during the pruning and quantization process. The proposed compression scheme is flexible to support supervised and unsupervised learning styles. Experiment results show GPUSQ-ViT scheme achieves state-of-the-art compression by reducing vision transformer models 6.4-12.7 times on model size and 30.3-62 times on FLOPs with negligible accuracy degradation on ImageNet classification, COCO detection and ADE20K segmentation benchmarking tasks. Moreover, GPUSQ-ViT can boost actual deployment performance by 1.39-1.79 times and 3.22-3.43 times of latency and throughput on A100 GPU, and 1.57-1.69 times and 2.11-2.51 times improvement of latency and throughput on AGX Orin.
In knowledge distillation, the knowledge from the teacher model is often too complex for the student model to thoroughly process. However, good teachers in real life always simplify complex material before teaching it to students. Inspired by this fact, we propose student-friendly knowledge distillation (SKD) to simplify teacher output into new knowledge representations, which makes the learning of the student model easier and more effective. SKD contains a softening processing and a learning simplifier. First, the softening processing uses the temperature hyperparameter to soften the output logits of the teacher model, which simplifies the output to some extent and makes it easier for the learning simplifier to process. The learning simplifier utilizes the attention mechanism to further simplify the knowledge of the teacher model and is jointly trained with the student model using the distillation loss, which means that the process of simplification is correlated with the training objective of the student model and ensures that the simplified new teacher knowledge representation is more suitable for the specific student model. Furthermore, since SKD does not change the form of the distillation loss, it can be easily combined with other distillation methods that are based on the logits or features of intermediate layers to enhance its effectiveness. Therefore, SKD has wide applicability. The experimental results on the CIFAR-100 and ImageNet datasets show that our method achieves state-of-the-art performance while maintaining high training efficiency.
Knowledge distillation has attracted a great deal of interest recently to compress pre-trained language models. However, existing knowledge distillation methods suffer from two limitations. First, the student model simply imitates the teacher's behavior while ignoring the underlying reasoning. Second, these methods usually focus on the transfer of sophisticated model-specific knowledge but overlook data-specific knowledge. In this paper, we present a novel attribution-driven knowledge distillation approach, which explores the token-level rationale behind the teacher model based on Integrated Gradients (IG) and transfers attribution knowledge to the student model. To enhance the knowledge transfer of model reasoning and generalization, we further explore multi-view attribution distillation on all potential decisions of the teacher. Comprehensive experiments are conducted with BERT on the GLUE benchmark. The experimental results demonstrate the superior performance of our approach to several state-of-the-art methods.
Large pre-trained language models (PLMs) have garnered significant attention for their versatility and potential for solving a wide spectrum of natural language processing (NLP) tasks. However, the cost of running these PLMs may be prohibitive. Furthermore, PLMs may not be open-sourced due to commercial considerations and potential risks of misuse, such as GPT-3. The parameters and gradients of PLMs are unavailable in this scenario. To solve the issue, black-box tuning has been proposed, which utilizes derivative-free optimization (DFO), instead of gradient descent, for training task-specific continuous prompts. However, these gradient-free methods still exhibit a significant gap compared to gradient-based methods. In this paper, we introduce gradient descent into black-box tuning scenario through knowledge distillation. Furthermore, we propose a novel method GDFO, which integrates gradient descent and derivative-free optimization to optimize task-specific continuous prompts in a harmonized manner. Experimental results show that GDFO can achieve significant performance gains over previous state-of-the-art methods.
In this work, we explore a Multilingual Information Retrieval (MLIR) task, where the collection includes documents in multiple languages. We demonstrate that applying state-of-the-art approaches developed for cross-lingual information retrieval to MLIR tasks leads to sub-optimal performance. This is due to the heterogeneous and imbalanced nature of multilingual collections -- some languages are better represented in the collection and some benefit from large-scale training data. To address this issue, we present KD-SPD, a novel soft prompt decoding approach for MLIR that implicitly "translates" the representation of documents in different languages into the same embedding space. To address the challenges of data scarcity and imbalance, we introduce a knowledge distillation strategy. The teacher model is trained on rich English retrieval data, and by leveraging bi-text data, our distillation framework transfers its retrieval knowledge to the multilingual document encoder. Therefore, our approach does not require any multilingual retrieval training data. Extensive experiments on three MLIR datasets with a total of 15 languages demonstrate that KD-SPD significantly outperforms competitive baselines in all cases. We conduct extensive analyses to show that our method has less language bias and better zero-shot transfer ability towards new languages.
It has been commonly observed that a teacher model with superior performance does not necessarily result in a stronger student, highlighting a discrepancy between current teacher training practices and effective knowledge transfer. In order to enhance the guidance of the teacher training process, we introduce the concept of distillation influence to determine the impact of distillation from each training sample on the student's generalization ability. In this paper, we propose Learning Good Teacher Matters (LGTM), an efficient training technique for incorporating distillation influence into the teacher's learning process. By prioritizing samples that are likely to enhance the student's generalization ability, our LGTM outperforms 10 common knowledge distillation baselines on 6 text classification tasks in the GLUE benchmark.
Knowledge Distillation (KD) is a predominant approach for BERT compression. Previous KD-based methods focus on designing extra alignment losses for the student model to mimic the behavior of the teacher model. These methods transfer the knowledge in an indirect way. In this paper, we propose a novel Weight-Inherited Distillation (WID), which directly transfers knowledge from the teacher. WID does not require any additional alignment loss and trains a compact student by inheriting the weights, showing a new perspective of knowledge distillation. Specifically, we design the row compactors and column compactors as mappings and then compress the weights via structural re-parameterization. Experimental results on the GLUE and SQuAD benchmarks show that WID outperforms previous state-of-the-art KD-based baselines. Further analysis indicates that WID can also learn the attention patterns from the teacher model without any alignment loss on attention distributions. The code is available at https://github.com/wutaiqiang/WID-NAACL2024.
Knowledge Distillation (KD) is a powerful technique for transferring knowledge between neural network models, where a pre-trained teacher model is used to facilitate the training of the target student model. However, the availability of a suitable teacher model is not always guaranteed. To address this challenge, Self-Knowledge Distillation (SKD) attempts to construct a teacher model from itself. Existing SKD methods add Auxiliary Classifiers (AC) to intermediate layers of the model or use the history models and models with different input data within the same class. However, these methods are computationally expensive and only capture time-wise and class-wise features of data. In this paper, we propose a lightweight SKD framework that utilizes multi-source information to construct a more informative teacher. Specifically, we introduce a Distillation with Reverse Guidance (DRG) method that considers different levels of information extracted by the model, including edge, shape, and detail of the input data, to construct a more informative teacher. Additionally, we design a Distillation with Shape-wise Regularization (DSR) method that ensures a consistent shape of ranked model output for all data. We validate the performance of the proposed DRG, DSR, and their combination through comprehensive experiments on various datasets and models. Our results demonstrate the superiority of the proposed methods over baselines (up to 2.87%) and state-of-the-art SKD methods (up to 1.15%), while being computationally efficient and robust. The code is available at https://github.com/xucong-parsifal/LightSKD.