Knowledge distillation (KD) is the process of transferring knowledge from a large model to a small one. It has gained increasing attention in the natural language processing community, driven by the demands of compressing ever-growing language models. In this work, we propose an f-DISTILL framework, which formulates sequence-level knowledge distillation as minimizing a generalized f-divergence function. We propose four distilling variants under our framework and show that existing SeqKD and ENGINE approaches are approximations of our f-DISTILL methods. We further derive step-wise decomposition for our f-DISTILL, reducing intractable sequence-level divergence to word-level losses that can be computed in a tractable manner. Experiments across four datasets show that our methods outperform existing KD approaches, and that our symmetric distilling losses can better force the student to learn from the teacher distribution.
Deep learning often faces the challenge of efficiently processing dynamic inputs, such as sensor data or user inputs. For example, an AI writing assistant is required to update its suggestions in real time as a document is edited. Re-running the model each time is expensive, even with compression techniques like knowledge distillation, pruning, or quantization. Instead, we take an incremental computing approach, looking to reuse calculations as the inputs change. However, the dense connectivity of conventional architectures poses a major obstacle to incremental computation, as even minor input changes cascade through the network and restrict information reuse. To address this, we use vector quantization to discretize intermediate values in the network, which filters out noisy and unnecessary modifications to hidden neurons, facilitating the reuse of their values. We apply this approach to the transformers architecture, creating an efficient incremental inference algorithm with complexity proportional to the fraction of the modified inputs. Our experiments with adapting the OPT-125M pre-trained language model demonstrate comparable accuracy on document classification while requiring 12.1X (median) fewer operations for processing sequences of atomic edits.
Waveform-based deep learning faces a dilemma between nonparametric and parametric approaches. On one hand, convolutional neural networks (convnets) may approximate any linear time-invariant system; yet, in practice, their frequency responses become more irregular as their receptive fields grow. On the other hand, a parametric model such as LEAF is guaranteed to yield Gabor filters, hence an optimal time-frequency localization; yet, this strong inductive bias comes at the detriment of representational capacity. In this paper, we aim to overcome this dilemma by introducing a neural audio model, named multiresolution neural network (MuReNN). The key idea behind MuReNN is to train separate convolutional operators over the octave subbands of a discrete wavelet transform (DWT). Since the scale of DWT atoms grows exponentially between octaves, the receptive fields of the subsequent learnable convolutions in MuReNN are dilated accordingly. For a given real-world dataset, we fit the magnitude response of MuReNN to that of a well-established auditory filterbank: Gammatone for speech, CQT for music, and third-octave for urban sounds, respectively. This is a form of knowledge distillation (KD), in which the filterbank ''teacher'' is engineered by domain knowledge while the neural network ''student'' is optimized from data. We compare MuReNN to the state of the art in terms of goodness of fit after KD on a hold-out set and in terms of Heisenberg time-frequency localization. Compared to convnets and Gabor convolutions, we find that MuReNN reaches state-of-the-art performance on all three optimization problems.
Machine learning makes multimedia data (e.g., images) more attractive, however, multimedia data is usually distributed and privacy sensitive. Multiple distributed multimedia clients can resort to federated learning (FL) to jointly learn a global shared model without requiring to share their private samples with any third-party entities. In this paper, we show that FL suffers from the cross-client generative adversarial networks (GANs)-based (C-GANs) attack, in which a malicious client (i.e., adversary) can reconstruct samples with the same distribution as the training samples from other clients (i.e., victims). Since a benign client's data can be leaked to the adversary, this attack brings the risk of local data leakage for clients in many security-critical FL applications. Thus, we propose Fed-EDKD (i.e., Federated Ensemble Data-free Knowledge Distillation) technique to improve the current popular FL schemes to resist C-GANs attack. In Fed-EDKD, each client submits a local model to the server for obtaining an ensemble global model. Then, to avoid model expansion, Fed-EDKD adopts data-free knowledge distillation techniques to transfer knowledge from the ensemble global model to a compressed model. By this way, Fed-EDKD reduces the adversary's control capability over the global model, so Fed-EDKD can effectively mitigate C-GANs attack. Finally, the experimental results demonstrate that Fed-EDKD significantly mitigates C-GANs attack while only incurring a slight accuracy degradation of FL.
Owing to the nature of privacy protection, federated recommender systems (FedRecs) have garnered increasing interest in the realm of on-device recommender systems. However, most existing FedRecs only allow participating clients to collaboratively train a recommendation model of the same public parameter size. Training a model of the same size for all clients can lead to suboptimal performance since clients possess varying resources. For example, clients with limited training data may prefer to train a smaller recommendation model to avoid excessive data consumption, while clients with sufficient data would benefit from a larger model to achieve higher recommendation accuracy. To address the above challenge, this paper introduces HeteFedRec, a novel FedRec framework that enables the assignment of personalized model sizes to participants. In HeteFedRec, we present a heterogeneous recommendation model aggregation strategy, including a unified dual-task learning mechanism and a dimensional decorrelation regularization, to allow knowledge aggregation among recommender models of different sizes. Additionally, a relation-based ensemble knowledge distillation method is proposed to effectively distil knowledge from heterogeneous item embeddings. Extensive experiments conducted on three real-world recommendation datasets demonstrate the effectiveness and efficiency of HeteFedRec in training federated recommender systems under heterogeneous settings.
Engineering knowledge-based (or expert) systems require extensive manual effort and domain knowledge. As Large Language Models (LLMs) are trained using an enormous amount of cross-domain knowledge, it becomes possible to automate such engineering processes. This paper presents an empirical automation and semi-automation framework for domain knowledge distillation using prompt engineering and the LLM ChatGPT. We assess the framework empirically in the autonomous driving domain and present our key observations. In our implementation, we construct the domain knowledge ontology by "chatting" with ChatGPT. The key finding is that while fully automated domain ontology construction is possible, human supervision and early intervention typically improve efficiency and output quality as they lessen the effects of response randomness and the butterfly effect. We, therefore, also develop a web-based distillation assistant enabling supervision and flexible intervention at runtime. We hope our findings and tools could inspire future research toward revolutionizing the engineering of knowledge-based systems across application domains.
Over the past few years, extensive research has been devoted to enhancing YOLO object detectors. Since its introduction, eight major versions of YOLO have been introduced with the purpose of improving its accuracy and efficiency. While the evident merits of YOLO have yielded to its extensive use in many areas, deploying it on resource-limited devices poses challenges. To address this issue, various neural network compression methods have been developed, which fall under three main categories, namely network pruning, quantization, and knowledge distillation. The fruitful outcomes of utilizing model compression methods, such as lowering memory usage and inference time, make them favorable, if not necessary, for deploying large neural networks on hardware-constrained edge devices. In this review paper, our focus is on pruning and quantization due to their comparative modularity. We categorize them and analyze the practical results of applying those methods to YOLOv5. By doing so, we identify gaps in adapting pruning and quantization for compressing YOLOv5, and provide future directions in this area for further exploration. Among several versions of YOLO, we specifically choose YOLOv5 for its excellent trade-off between recency and popularity in literature. This is the first specific review paper that surveys pruning and quantization methods from an implementation point of view on YOLOv5. Our study is also extendable to newer versions of YOLO as implementing them on resource-limited devices poses the same challenges that persist even today. This paper targets those interested in the practical deployment of model compression methods on YOLOv5, and in exploring different compression techniques that can be used for subsequent versions of YOLO.
Contrastive Language-Image Pre-training (CLIP) has become a promising language-supervised visual pre-training framework. This paper aims to distill small CLIP models supervised by a large teacher CLIP model. We propose several distillation strategies, including relation, feature, gradient and contrastive paradigms, to examine the effectiveness of CLIP-Knowledge Distillation (KD). We show that a simple feature mimicry with Mean Squared Error loss works surprisingly well. Moreover, interactive contrastive learning across teacher and student encoders is also effective in performance improvement. We explain that the success of CLIP-KD can be attributed to maximizing the feature similarity between teacher and student. The unified method is applied to distill several student models trained on CC3M+12M. CLIP-KD improves student CLIP models consistently over zero-shot ImageNet classification and cross-modal retrieval benchmarks. When using ViT-L/14 pretrained on Laion-400M as the teacher, CLIP-KD achieves 57.5\% and 55.4\% zero-shot top-1 ImageNet accuracy over ViT-B/16 and ResNet-50, surpassing the original CLIP without KD by 20.5\% and 20.1\% margins, respectively. Our code is released on https://github.com/winycg/CLIP-KD.
In this paper, we strive to answer the question "how to collaboratively learn convolutional neural network (CNN)-based and vision transformer (ViT)-based models by selecting and exchanging the reliable knowledge between them for semantic segmentation?" Accordingly, we propose an online knowledge distillation (KD) framework that can simultaneously learn compact yet effective CNN-based and ViT-based models with two key technical breakthroughs to take full advantage of CNNs and ViT while compensating their limitations. Firstly, we propose heterogeneous feature distillation (HFD) to improve students' consistency in low-layer feature space by mimicking heterogeneous features between CNNs and ViT. Secondly, to facilitate the two students to learn reliable knowledge from each other, we propose bidirectional selective distillation (BSD) that can dynamically transfer selective knowledge. This is achieved by 1) region-wise BSD determining the directions of knowledge transferred between the corresponding regions in the feature space and 2) pixel-wise BSD discerning which of the prediction knowledge to be transferred in the logit space. Extensive experiments on three benchmark datasets demonstrate that our proposed framework outperforms the state-of-the-art online distillation methods by a large margin, and shows its efficacy in learning collaboratively between ViT-based and CNN-based models.
Knowledge distillation aims to learn a lightweight student network from a pre-trained teacher network. In practice, existing knowledge distillation methods are usually infeasible when the original training data is unavailable due to some privacy issues and data management considerations. Therefore, data-free knowledge distillation approaches proposed to collect training instances from the Internet. However, most of them have ignored the common distribution shift between the instances from original training data and webly collected data, affecting the reliability of the trained student network. To solve this problem, we propose a novel method dubbed ``Knowledge Distillation between Different Distributions" (KD$^{3}$), which consists of three components. Specifically, we first dynamically select useful training instances from the webly collected data according to the combined predictions of teacher network and student network. Subsequently, we align both the weighted features and classifier parameters of the two networks for knowledge memorization. Meanwhile, we also build a new contrastive learning block called MixDistribution to generate perturbed data with a new distribution for instance alignment, so that the student network can further learn a distribution-invariant representation. Intensive experiments on various benchmark datasets demonstrate that our proposed KD$^{3}$ can outperform the state-of-the-art data-free knowledge distillation approaches.
Sampling from diffusion probabilistic models (DPMs) can be viewed as a piecewise distribution transformation, which generally requires hundreds or thousands of steps of the inverse diffusion trajectory to get a high-quality image. Recent progress in designing fast samplers for DPMs achieves a trade-off between sampling speed and sample quality by knowledge distillation or adjusting the variance schedule or the denoising equation. However, it can't be optimal in both aspects and often suffer from mode mixture in short steps. To tackle this problem, we innovatively regard inverse diffusion as an optimal transport (OT) problem between latents at different stages and propose the DPM-OT, a unified learning framework for fast DPMs with a direct expressway represented by OT map, which can generate high-quality samples within around 10 function evaluations. By calculating the semi-discrete optimal transport map between the data latents and the white noise, we obtain an expressway from the prior distribution to the data distribution, while significantly alleviating the problem of mode mixture. In addition, we give the error bound of the proposed method, which theoretically guarantees the stability of the algorithm. Extensive experiments validate the effectiveness and advantages of DPM-OT in terms of speed and quality (FID and mode mixture), thus representing an efficient solution for generative modeling. Source codes are available at https://github.com/cognaclee/DPM-OT
Despite the empirical success and practical significance of (relational) knowledge distillation that matches (the relations of) features between teacher and student models, the corresponding theoretical interpretations remain limited for various knowledge distillation paradigms. In this work, we take an initial step toward a theoretical understanding of relational knowledge distillation (RKD), with a focus on semi-supervised classification problems. We start by casting RKD as spectral clustering on a population-induced graph unveiled by a teacher model. Via a notion of clustering error that quantifies the discrepancy between the predicted and ground truth clusterings, we illustrate that RKD over the population provably leads to low clustering error. Moreover, we provide a sample complexity bound for RKD with limited unlabeled samples. For semi-supervised learning, we further demonstrate the label efficiency of RKD through a general framework of cluster-aware semi-supervised learning that assumes low clustering errors. Finally, by unifying data augmentation consistency regularization into this cluster-aware framework, we show that despite the common effect of learning accurate clusterings, RKD facilitates a "global" perspective through spectral clustering, whereas consistency regularization focuses on a "local" perspective via expansion.
Retinal image matching plays a crucial role in monitoring disease progression and treatment response. However, datasets with matched keypoints between temporally separated pairs of images are not available in abundance to train transformer-based model. We propose a novel approach based on reverse knowledge distillation to train large models with limited data while preventing overfitting. Firstly, we propose architectural modifications to a CNN-based semi-supervised method called SuperRetina that help us improve its results on a publicly available dataset. Then, we train a computationally heavier model based on a vision transformer encoder using the lighter CNN-based model, which is counter-intuitive in the field knowledge-distillation research where training lighter models based on heavier ones is the norm. Surprisingly, such reverse knowledge distillation improves generalization even further. Our experiments suggest that high-dimensional fitting in representation space may prevent overfitting unlike training directly to match the final output. We also provide a public dataset with annotations for retinal image keypoint detection and matching to help the research community develop algorithms for retinal image applications.
Neural network quantization aims to accelerate and trim full-precision neural network models by using low bit approximations. Methods adopting the quantization aware training (QAT) paradigm have recently seen a rapid growth, but are often conceptually complicated. This paper proposes a novel and highly effective QAT method, quantized feature distillation (QFD). QFD first trains a quantized (or binarized) representation as the teacher, then quantize the network using knowledge distillation (KD). Quantitative results show that QFD is more flexible and effective (i.e., quantization friendly) than previous quantization methods. QFD surpasses existing methods by a noticeable margin on not only image classification but also object detection, albeit being much simpler. Furthermore, QFD quantizes ViT and Swin-Transformer on MS-COCO detection and segmentation, which verifies its potential in real world deployment. To the best of our knowledge, this is the first time that vision transformers have been quantized in object detection and image segmentation tasks.
Movement paths are used widely in intelligent transportation and smart city applications. To serve such applications, path representation learning aims to provide compact representations of paths that enable efficient and accurate operations when used for different downstream tasks such as path ranking and travel cost estimation. In many cases, it is attractive that the path representation learning is lightweight and scalable; in resource-limited environments and under green computing limitations, it is essential. Yet, existing path representation learning studies focus on accuracy and pay at most secondary attention to resource consumption and scalability. We propose a lightweight and scalable path representation learning framework, termed LightPath, that aims to reduce resource consumption and achieve scalability without affecting accuracy, thus enabling broader applicability. More specifically, we first propose a sparse auto-encoder that ensures that the framework achieves good scalability with respect to path length. Next, we propose a relational reasoning framework to enable faster training of more robust sparse path encoders. We also propose global-local knowledge distillation to further reduce the size and improve the performance of sparse path encoders. Finally, we report extensive experiments on two real-world datasets to offer insight into the efficiency, scalability, and effectiveness of the proposed framework.
Multi-choice Machine Reading Comprehension (MRC) is a challenging extension of Natural Language Processing (NLP) that requires the ability to comprehend the semantics and logical relationships between entities in a given text. The MRC task has traditionally been viewed as a process of answering questions based on the given text. This single-stage approach has often led the network to concentrate on generating the correct answer, potentially neglecting the comprehension of the text itself. As a result, many prevalent models have faced challenges in performing well on this task when dealing with longer texts. In this paper, we propose a two-stage knowledge distillation method that teaches the model to better comprehend the document by dividing the MRC task into two separate stages. Our experimental results show that the student model, when equipped with our method, achieves significant improvements, demonstrating the effectiveness of our method.
We present DARTS, a transformer model for reference-based image super-resolution. DARTS learns joint representations of two image distributions to enhance the content of low-resolution input images through matching correspondences learned from high-resolution reference images. Current state-of-the-art techniques in reference-based image super-resolution are based on a multi-network, multi-stage architecture. In this work, we adapt the double attention block from the GAN literature, processing the two visual streams separately and combining self-attention and cross-attention blocks through a gating attention strategy. Our work demonstrates how the attention mechanism can be adapted for the particular requirements of reference-based image super-resolution, significantly simplifying the architecture and training pipeline. We show that our transformer-based model performs competitively with state-of-the-art models, while maintaining a simpler overall architecture and training process. In particular, we obtain state-of-the-art on the SUN80 dataset, with a PSNR/SSIM of 29.83 / .809. These results show that attention alone is sufficient for the RSR task, without multiple purpose-built subnetworks, knowledge distillation, or multi-stage training.
We propose an efficient modeling framework for cross-lingual named entity recognition in semi-structured text data. Our approach relies on both knowledge distillation and consistency training. The modeling framework leverages knowledge from a large language model (XLMRoBERTa) pre-trained on the source language, with a student-teacher relationship (knowledge distillation). The student model incorporates unsupervised consistency training (with KL divergence loss) on the low-resource target language. We employ two independent datasets of SMSs in English and Arabic, each carrying semi-structured banking transaction information, and focus on exhibiting the transfer of knowledge from English to Arabic. With access to only 30 labeled samples, our model can generalize the recognition of merchants, amounts, and other fields from English to Arabic. We show that our modeling approach, while efficient, performs best overall when compared to state-of-the-art approaches like DistilBERT pre-trained on the target language or a supervised model directly trained on labeled data in the target language. Our experiments show that it is enough to learn to recognize entities in English to reach reasonable performance in a low-resource language in the presence of a few labeled samples of semi-structured data. The proposed framework has implications for developing multi-lingual applications, especially in geographies where digital endeavors rely on both English and one or more low-resource language(s), sometimes mixed with English or employed singly.
Knowledge distillation, a well-known model compression technique, is an active research area in both computer vision and remote sensing communities. In this paper, we evaluate in a remote sensing context various off-the-shelf object detection knowledge distillation methods which have been originally developed on generic computer vision datasets such as Pascal VOC. In particular, methods covering both logit mimicking and feature imitation approaches are applied for vehicle detection using the well-known benchmarks such as xView and VEDAI datasets. Extensive experiments are performed to compare the relative performance and interrelationships of the methods. Experimental results show high variations and confirm the importance of result aggregation and cross validation on remote sensing datasets.
As the most fundamental scene understanding tasks, object detection and segmentation have made tremendous progress in deep learning era. Due to the expensive manual labeling cost, the annotated categories in existing datasets are often small-scale and pre-defined, i.e., state-of-the-art fully-supervised detectors and segmentors fail to generalize beyond the closed vocabulary. To resolve this limitation, in the last few years, the community has witnessed an increasing attention toward Open-Vocabulary Detection (OVD) and Segmentation (OVS). By ``open-vocabulary'', we mean that the models can classify objects beyond pre-defined categories. In this survey, we provide a comprehensive review on recent developments of OVD and OVS. A taxonomy is first developed to organize different tasks and methodologies. We find that the permission and usage of weak supervision signals can well discriminate different methodologies, including: visual-semantic space mapping, novel visual feature synthesis, region-aware training, pseudo-labeling, knowledge distillation, and transfer learning. The proposed taxonomy is universal across different tasks, covering object detection, semantic/instance/panoptic segmentation, 3D and video understanding. The main design principles, key challenges, development routes, methodology strengths, and weaknesses are thoroughly analyzed. In addition, we benchmark each task along with the vital components of each method in appendix and updated online at https://github.com/seanzhuh/awesome-open-vocabulary-detection-and-segmentation. Finally, several promising directions are provided and discussed to stimulate future research.