Knowledge distillation is a popular approach for enhancing the performance of ''student'' models, with lower representational capacity, by taking advantage of more powerful ''teacher'' models. Despite its apparent simplicity and widespread use, the underlying mechanics behind knowledge distillation (KD) are still not fully understood. In this work, we shed new light on the inner workings of this method, by examining it from an optimization perspective. We show that, in the context of linear and deep linear models, KD can be interpreted as a novel type of stochastic variance reduction mechanism. We provide a detailed convergence analysis of the resulting dynamics, which hold under standard assumptions for both strongly-convex and non-convex losses, showing that KD acts as a form of partial variance reduction, which can reduce the stochastic gradient noise, but may not eliminate it completely, depending on the properties of the ''teacher'' model. Our analysis puts further emphasis on the need for careful parametrization of KD, in particular w.r.t. the weighting of the distillation loss, and is validated empirically on both linear models and deep neural networks.
Representational transfer from publicly available models is a promising technique for improving medical image classification, especially in long-tailed datasets with rare diseases. However, existing methods often overlook the frequency-dependent behavior of these models, thereby limiting their effectiveness in transferring representations and generalizations to rare diseases. In this paper, we propose FoPro-KD, a novel framework that leverages the power of frequency patterns learned from frozen pre-trained models to enhance their transferability and compression, presenting a few unique insights: 1) We demonstrate that leveraging representations from publicly available pre-trained models can substantially improve performance, specifically for rare classes, even when utilizing representations from a smaller pre-trained model. 2) We observe that pre-trained models exhibit frequency preferences, which we explore using our proposed Fourier Prompt Generator (FPG), allowing us to manipulate specific frequencies in the input image, enhancing the discriminative representational transfer. 3) By amplifying or diminishing these frequencies in the input image, we enable Effective Knowledge Distillation (EKD). EKD facilitates the transfer of knowledge from pre-trained models to smaller models. Through extensive experiments in long-tailed gastrointestinal image recognition and skin lesion classification, where rare diseases are prevalent, our FoPro-KD framework outperforms existing methods, enabling more accessible medical models for rare disease classification. Code is available at https://github.com/xmed-lab/FoPro-KD.
Distillation from Weak Teacher (DWT) is a method of transferring knowledge from a smaller, weaker teacher model to a larger student model to improve its performance. Previous studies have shown that DWT can be effective in the vision domain and natural language processing (NLP) pre-training stage. Specifically, DWT shows promise in practical scenarios, such as enhancing new generation or larger models using pre-trained yet older or smaller models and lacking a resource budget. However, the optimal conditions for using DWT have yet to be fully investigated in NLP pre-training. Therefore, this study examines three key factors to optimize DWT, distinct from those used in the vision domain or traditional knowledge distillation. These factors are: (i) the impact of teacher model quality on DWT effectiveness, (ii) guidelines for adjusting the weighting value for DWT loss, and (iii) the impact of parameter remapping as a student model initialization technique for DWT.
Computational Pathology (CPATH) systems have the potential to automate diagnostic tasks. However, the artifacts on the digitized histological glass slides, known as Whole Slide Images (WSIs), may hamper the overall performance of CPATH systems. Deep Learning (DL) models such as Vision Transformers (ViTs) may detect and exclude artifacts before running the diagnostic algorithm. A simple way to develop robust and generalized ViTs is to train them on massive datasets. Unfortunately, acquiring large medical datasets is expensive and inconvenient, prompting the need for a generalized artifact detection method for WSIs. In this paper, we present a student-teacher recipe to improve the classification performance of ViT for the air bubbles detection task. ViT, trained under the student-teacher framework, boosts its performance by distilling existing knowledge from the high-capacity teacher model. Our best-performing ViT yields 0.961 and 0.911 F1-score and MCC, respectively, observing a 7% gain in MCC against stand-alone training. The proposed method presents a new perspective of leveraging knowledge distillation over transfer learning to encourage the use of customized transformers for efficient preprocessing pipelines in the CPATH systems.
Knowledge distillation (KD) exploits a large well-trained model (i.e., teacher) to train a small student model on the same dataset for the same task. Treating teacher features as knowledge, prevailing methods of knowledge distillation train student by aligning its features with the teacher's, e.g., by minimizing the KL-divergence between their logits or L2 distance between their intermediate features. While it is natural to believe that better alignment of student features to the teacher better distills teacher knowledge, simply forcing this alignment does not directly contribute to the student's performance, e.g., classification accuracy. In this work, we propose to align student features with class-mean of teacher features, where class-mean naturally serves as a strong classifier. To this end, we explore baseline techniques such as adopting the cosine distance based loss to encourage the similarity between student features and their corresponding class-means of the teacher. Moreover, we train the student to produce large-norm features, inspired by other lines of work (e.g., model pruning and domain adaptation), which find the large-norm features to be more significant. Finally, we propose a rather simple loss term (dubbed ND loss) to simultaneously (1) encourage student to produce large-\emph{norm} features, and (2) align the \emph{direction} of student features and teacher class-means. Experiments on standard benchmarks demonstrate that our explored techniques help existing KD methods achieve better performance, i.e., higher classification accuracy on ImageNet and CIFAR100 datasets, and higher detection precision on COCO dataset. Importantly, our proposed ND loss helps the most, leading to the state-of-the-art performance on these benchmarks. The source code is available at \url{https://github.com/WangYZ1608/Knowledge-Distillation-via-ND}.
Multilingual neural machine translation has witnessed remarkable progress in recent years. However, the long-tailed distribution of multilingual corpora poses a challenge of Pareto optimization, i.e., optimizing for some languages may come at the cost of degrading the performance of others. Existing balancing training strategies are equivalent to a series of Pareto optimal solutions, which trade off on a Pareto frontier. In this work, we propose a new training framework, Pareto Mutual Distillation (Pareto-MD), towards pushing the Pareto frontier outwards rather than making trade-offs. Specifically, Pareto-MD collaboratively trains two Pareto optimal solutions that favor different languages and allows them to learn from the strengths of each other via knowledge distillation. Furthermore, we introduce a novel strategy to enable stronger communication between Pareto optimal solutions and broaden the applicability of our approach. Experimental results on the widely-used WMT and TED datasets show that our method significantly pushes the Pareto frontier and outperforms baselines by up to +2.46 BLEU.
Object Re-identification (ReID) aims to retrieve the probe object from many gallery images with the ReID model inferred based on a stationary camera-free dataset by associating and collecting the identities across all camera views. When deploying the ReID algorithm in real-world scenarios, the aspect of storage, privacy constraints, and dynamic changes of cameras would degrade its generalizability and applicability. Treating each camera's data independently, we introduce a novel ReID task named Camera-Incremental Object Re-identification (CIOR) by continually optimizing the ReID mode from the incoming stream of the camera dataset. Since the identities under different camera views might describe the same object, associating and distilling the knowledge of common identities would boost the discrimination and benefit from alleviating the catastrophic forgetting. In this paper, we propose a novel Identity Knowledge Evolution (IKE) framework for CIOR, consisting of the Identity Knowledge Association (IKA), Identity Knowledge Distillation (IKD), and Identity Knowledge Update (IKU). IKA is proposed to discover the common identities between the current identity and historical identities. IKD has applied to distillate historical identity knowledge from common identities and quickly adapt the historical model to the current camera view. After each camera has been trained, IKU is applied to continually expand the identity knowledge by combining the historical and current identity memories. The evaluation of Market-CL and Veri-CL shows the Identity Knowledge Evolution (IKE) effectiveness for CIOR. code:https://github.com/htyao89/Camera-Incremental-Object-ReID
The tremendous success of large models trained on extensive datasets demonstrates that scale is a key ingredient in achieving superior results. Therefore, the reflection on the rationality of designing knowledge distillation (KD) approaches for limited-capacity architectures solely based on small-scale datasets is now deemed imperative. In this paper, we identify the \emph{small data pitfall} that presents in previous KD methods, which results in the underestimation of the power of vanilla KD framework on large-scale datasets such as ImageNet-1K. Specifically, we show that employing stronger data augmentation techniques and using larger datasets can directly decrease the gap between vanilla KD and other meticulously designed KD variants. This highlights the necessity of designing and evaluating KD approaches in the context of practical scenarios, casting off the limitations of small-scale datasets. Our investigation of the vanilla KD and its variants in more complex schemes, including stronger training strategies and different model capacities, demonstrates that vanilla KD is elegantly simple but astonishingly effective in large-scale scenarios. Without bells and whistles, we obtain state-of-the-art ResNet-50, ViT-S, and ConvNeXtV2-T models for ImageNet, which achieve 83.1\%, 84.3\%, and 85.0\% top-1 accuracy, respectively. PyTorch code and checkpoints can be found at https://github.com/Hao840/vanillaKD.
Semantic occupancy prediction aims to infer dense geometry and semantics of surroundings for an autonomous agent to operate safely in the 3D environment. Existing occupancy prediction methods are almost entirely trained on human-annotated volumetric data. Although of high quality, the generation of such 3D annotations is laborious and costly, restricting them to a few specific object categories in the training dataset. To address this limitation, this paper proposes Open Vocabulary Occupancy (OVO), a novel approach that allows semantic occupancy prediction of arbitrary classes but without the need for 3D annotations during training. Keys to our approach are (1) knowledge distillation from a pre-trained 2D open-vocabulary segmentation model to the 3D occupancy network, and (2) pixel-voxel filtering for high-quality training data generation. The resulting framework is simple, compact, and compatible with most state-of-the-art semantic occupancy prediction models. On NYUv2 and SemanticKITTI datasets, OVO achieves competitive performance compared to supervised semantic occupancy prediction approaches. Furthermore, we conduct extensive analyses and ablation studies to offer insights into the design of the proposed framework. Our code is publicly available at https://github.com/dzcgaara/OVO.
Cross-lingual named entity recognition (NER) aims to train an NER system that generalizes well to a target language by leveraging labeled data in a given source language. Previous work alleviates the data scarcity problem by translating source-language labeled data or performing knowledge distillation on target-language unlabeled data. However, these methods may suffer from label noise due to the automatic labeling process. In this paper, we propose CoLaDa, a Collaborative Label Denoising Framework, to address this problem. Specifically, we first explore a model-collaboration-based denoising scheme that enables models trained on different data sources to collaboratively denoise pseudo labels used by each other. We then present an instance-collaboration-based strategy that considers the label consistency of each token's neighborhood in the representation space for denoising. Experiments on different benchmark datasets show that the proposed CoLaDa achieves superior results compared to previous methods, especially when generalizing to distant languages.
Recently, various intermediate layer distillation (ILD) objectives have been shown to improve compression of BERT models via Knowledge Distillation (KD). However, a comprehensive evaluation of the objectives in both task-specific and task-agnostic settings is lacking. To the best of our knowledge, this is the first work comprehensively evaluating distillation objectives in both settings. We show that attention transfer gives the best performance overall. We also study the impact of layer choice when initializing the student from the teacher layers, finding a significant impact on the performance in task-specific distillation. For vanilla KD and hidden states transfer, initialisation with lower layers of the teacher gives a considerable improvement over higher layers, especially on the task of QNLI (up to an absolute percentage change of 17.8 in accuracy). Attention transfer behaves consistently under different initialisation settings. We release our code as an efficient transformer-based model distillation framework for further studies.
As language models increase in size by the day, methods for efficient inference are critical to leveraging their capabilities for various applications. Prior work has investigated techniques like model pruning, knowledge distillation, and data multiplexing to increase model throughput without sacrificing accuracy. In this paper, we combine two such methods -- structured pruning and data multiplexing -- to compound the speedup gains obtained by either method. Our approach, PruMUX, obtains up to 7.5-29.5X throughput improvement over BERT-base model with accuracy threshold from 80% to 74%. We further study various combinations of parameters (such as sparsity and multiplexing factor) in the two techniques to provide a comprehensive analysis of the tradeoff between accuracy and throughput in the resulting models. We then propose Auto-PruMUX, a meta-level model that can predict the high-performance parameters for pruning and multiplexing given a desired accuracy loss budget, providing a practical method to leverage the combination effectively.
Self-supervised learning general-purpose audio representations have demonstrated high performance in a variety of tasks. Although they can be optimized for application by fine-tuning, even higher performance can be expected if they can be specialized to pre-train for an application. This paper explores the challenges and solutions in specializing general-purpose audio representations for a specific application using speech, a highly demanding field, as an example. We enhance Masked Modeling Duo (M2D), a general-purpose model, to close the performance gap with state-of-the-art (SOTA) speech models. To do so, we propose a new task, denoising distillation, to learn from fine-grained clustered features, and M2D for Speech (M2D-S), which jointly learns the denoising distillation task and M2D masked prediction task. Experimental results show that M2D-S performs comparably to or outperforms SOTA speech models on the SUPERB benchmark, demonstrating that M2D can specialize in a demanding field. Our code is available at: https://github.com/nttcslab/m2d/tree/master/speech
Existing feature distillation methods commonly adopt the One-to-one Representation Matching between any pre-selected teacher-student layer pair. In this paper, we present N-to-One Representation (NORM), a new two-stage knowledge distillation method, which relies on a simple Feature Transform (FT) module consisting of two linear layers. In view of preserving the intact information learnt by the teacher network, during training, our FT module is merely inserted after the last convolutional layer of the student network. The first linear layer projects the student representation to a feature space having N times feature channels than the teacher representation from the last convolutional layer, and the second linear layer contracts the expanded output back to the original feature space. By sequentially splitting the expanded student representation into N non-overlapping feature segments having the same number of feature channels as the teacher's, they can be readily forced to approximate the intact teacher representation simultaneously, formulating a novel many-to-one representation matching mechanism conditioned on a single teacher-student layer pair. After training, such an FT module will be naturally merged into the subsequent fully connected layer thanks to its linear property, introducing no extra parameters or architectural modifications to the student network at inference. Extensive experiments on different visual recognition benchmarks demonstrate the leading performance of our method. For instance, the ResNet18|MobileNet|ResNet50-1/4 model trained by NORM reaches 72.14%|74.26%|68.03% top-1 accuracy on the ImageNet dataset when using a pre-trained ResNet34|ResNet50|ResNet50 model as the teacher, achieving an absolute improvement of 2.01%|4.63%|3.03% against the individually trained counterpart. Code is available at https://github.com/OSVAI/NORM
In this paper, we delve deeper into the Kullback-Leibler (KL) Divergence loss and mathematically prove that it is equivalent to the Decoupled Kullback-Leibler (DKL) Divergence loss that consists of 1) a weighted Mean Square Error (wMSE) loss and 2) a Cross-Entropy loss incorporating soft labels. Thanks to the decomposed formulation of DKL loss, we have identified two areas for improvement. Firstly, we address the limitation of KL/DKL in scenarios like knowledge distillation by breaking its asymmetric optimization property. This modification ensures that the $\mathbf{w}$MSE component is always effective during training, providing extra constructive cues. Secondly, we introduce class-wise global information into KL/DKL to mitigate bias from individual samples. With these two enhancements, we derive the Improved Kullback-Leibler (IKL) Divergence loss and evaluate its effectiveness by conducting experiments on CIFAR-10/100 and ImageNet datasets, focusing on adversarial training, and knowledge distillation tasks. The proposed approach achieves new state-of-the-art adversarial robustness on the public leaderboard -- RobustBench and competitive performance on knowledge distillation, demonstrating the substantial practical merits. Our code is available at https://github.com/jiequancui/DKL.
Many text mining models are constructed by fine-tuning a large deep pre-trained language model (PLM) in downstream tasks. However, a significant challenge nowadays is maintaining performance when we use a lightweight model with limited labelled samples. We present DisCo, a semi-supervised learning (SSL) framework for fine-tuning a cohort of small student models generated from a large PLM using knowledge distillation. Our key insight is to share complementary knowledge among distilled student cohorts to promote their SSL effectiveness. DisCo employs a novel co-training technique to optimize a cohort of multiple small student models by promoting knowledge sharing among students under diversified views: model views produced by different distillation strategies and data views produced by various input augmentations. We evaluate DisCo on both semi-supervised text classification and extractive summarization tasks. Experimental results show that DisCo can produce student models that are 7.6 times smaller and 4.8 times faster in inference than the baseline PLMs while maintaining comparable performance. We also show that DisCo-generated student models outperform the similar-sized models elaborately tuned in distinct tasks.
The practice of transferring knowledge from a sophisticated, proprietary large language model (LLM) to a compact, open-source LLM has garnered considerable attention. Previous works have focused on a unidirectional knowledge distillation way by aligning the responses of the student model with those of the teacher model to a set of instructions. Nevertheless, they overlooked the possibility of incorporating any reciprocal "feedback"--identifying challenging instructions where the student model's performance falls short--to boost the student model's proficiency iteratively. To this end, we propose a novel adversarial distillation framework for a more efficient knowledge transfer. Leveraging the versatile role adaptability of LLMs, we prompt the teacher model to identify "hard" instructions and generate new "hard" instructions for the student model, creating a three-stage adversarial loop of imitation, discrimination, and generation. By applying this adversarial framework, we successfully transfer knowledge from ChatGPT to a student model (named Lion), using a mere 70k training data. Our results show that Lion-13B not only achieves comparable open-ended generation capabilities to ChatGPT but surpasses conventional state-of-the-art (SOTA) instruction-tuned models like Vicuna-13B by 55.4% in challenging zero-shot reasoning benchmarks such as BIG-Bench Hard (BBH) and 16.7% on AGIEval. Code and model can be found at https://github.com/YJiangcm/Lion.
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$95\% GLUE score of the teacher.
Diffusion models have recently achieved astonishing performance in generating high-fidelity photo-realistic images. Given their huge success, it is still unclear whether synthetic images are applicable for knowledge distillation when real images are unavailable. In this paper, we extensively study whether and how synthetic images produced from state-of-the-art diffusion models can be used for knowledge distillation without access to real images, and obtain three key conclusions: (1) synthetic data from diffusion models can easily lead to state-of-the-art performance among existing synthesis-based distillation methods, (2) low-fidelity synthetic images are better teaching materials, and (3) relatively weak classifiers are better teachers. Code is available at https://github.com/zhengli97/DM-KD.
Many-to-many multimodal summarization (M$^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.