Since pioneering work of Hinton et al., knowledge distillation based on Kullback-Leibler Divergence (KL-Div) has been predominant, and recently its variants have achieved compelling performance. However, KL-Div only compares probabilities of the corresponding category between the teacher and student while lacking a mechanism for cross-category comparison. Besides, KL-Div is problematic when applied to intermediate layers, as it cannot handle non-overlapping distributions and is unaware of geometry of the underlying manifold. To address these downsides, we propose a methodology of Wasserstein Distance (WD) based knowledge distillation. Specifically, we propose a logit distillation method called WKD-L based on discrete WD, which performs cross-category comparison of probabilities and thus can explicitly leverage rich interrelations among categories. Moreover, we introduce a feature distillation method called WKD-F, which uses a parametric method for modeling feature distributions and adopts continuous WD for transferring knowledge from intermediate layers. Comprehensive evaluations on image classification and object detection have shown (1) for logit distillation WKD-L outperforms very strong KL-Div variants; (2) for feature distillation WKD-F is superior to the KL-Div counterparts and state-of-the-art competitors. The source code is available at https://peihuali.org/WKD
Pruning can be an effective method of compressing large pre-trained models for inference speed acceleration. Previous pruning approaches rely on access to the original training dataset for both pruning and subsequent fine-tuning. However, access to the training data can be limited due to concerns such as data privacy and commercial confidentiality. Furthermore, with covariate shift (disparities between test and training data distributions), pruning and finetuning with training datasets can hinder the generalization of the pruned model to test data. To address these issues, pruning and finetuning the model with test time samples becomes essential. However, test-time model pruning and fine-tuning incur additional computation costs and slow down the model's prediction speed, thus posing efficiency issues. Existing pruning methods are not efficient enough for test time model pruning setting, since finetuning the pruned model is needed to evaluate the importance of removable components. To address this, we propose two variables to approximate the fine-tuned accuracy. We then introduce an efficient pruning method that considers the approximated finetuned accuracy and potential inference latency saving. To enhance fine-tuning efficiency, we propose an efficient knowledge distillation method that only needs to generate pseudo labels for a small set of finetuning samples one time, thereby reducing the expensive pseudo-label generation cost. Experimental results demonstrate that our method achieves a comparable or superior tradeoff between test accuracy and inference latency, with a 32% relative reduction in pruning and finetuning time compared to the best existing method.
Knowledge Distillation (KD) is a promising approach for unsupervised Anomaly Detection (AD). However, the student network's over-generalization often diminishes the crucial representation differences between teacher and student in anomalous regions, leading to detection failures. To addresses this problem, the widely accepted Reverse Distillation (RD) paradigm designs the asymmetry teacher and student, using an encoder as teacher and a decoder as student. Yet, the design of RD does not ensure that the teacher encoder effectively distinguishes between normal and abnormal features or that the student decoder generates anomaly-free features. Additionally, the absence of skip connections results in a loss of fine details during feature reconstruction. To address these issues, we propose RD with Expert, which introduces a novel Expert-Teacher-Student network for simultaneous distillation of both the teacher encoder and student decoder. The added expert network enhances the student's ability to generate normal features and optimizes the teacher's differentiation between normal and abnormal features, reducing missed detections. Additionally, Guided Information Injection is designed to filter and transfer features from teacher to student, improving detail reconstruction and minimizing false positives. Experiments on several benchmarks prove that our method outperforms existing unsupervised AD methods under RD paradigm, fully unlocking RD's potential.
Multimodal multihop question answering (MMQA) requires reasoning over images and text from multiple sources. Despite advances in visual question answering, this multihop setting remains underexplored due to a lack of quality datasets. Existing methods focus on single-hop, single-modality, or short texts, limiting real-world applications like interpreting educational documents with long, multimodal content. To fill this gap, we introduce FM2DS, the first framework for creating a high-quality dataset for MMQA. Our approach consists of a 5-stage pipeline that involves acquiring relevant multimodal documents from Wikipedia, synthetically generating high-level questions and answers, and validating them through rigorous criteria to ensure data quality. We evaluate our methodology by training models on our synthesized dataset and testing on two benchmarks: MultimodalQA and WebQA. Our results demonstrate that, with an equal sample size, models trained on our synthesized data outperform those trained on human-collected data by 1.9 in exact match (EM) score on average. Additionally, we introduce M2QA-Bench with 1k samples, the first benchmark for MMQA on long documents, generated using FM2DS and refined by human annotators. We believe our data synthesis method will serve as a strong foundation for training and evaluating MMQA models.
In this work, we compare the domain-specific translation performance of open-source autoregressive decoder-only large language models (LLMs) with task-oriented machine translation (MT) models. Our experiments focus on the medical domain and cover four language directions with varied resource availability: English-to-French, English-to-Portuguese, English-to-Swahili, and Swahili-to-English. Despite recent advancements, LLMs demonstrate a significant quality gap in specialized translation compared to multilingual encoder-decoder MT models such as NLLB-200. Our results indicate that NLLB-200 3.3B outperforms all evaluated LLMs in the 7-8B parameter range across three out of the four language directions. While fine-tuning improves the performance of LLMs such as Mistral and Llama, these models still underperform compared to fine-tuned NLLB-200 3.3B models. Our findings highlight the ongoing need for specialized MT models to achieve high-quality domain-specific translation, especially in medium-resource and low-resource settings. Moreover, the superior performance of larger LLMs over their 8B variants suggests potential value in pre-training domain-specific medium-sized language models, employing targeted data selection and knowledge distillation approaches to enhance both quality and efficiency in specialized translation tasks.
Conventional methods for PAN-sharpening often struggle to restore fine details due to limitations in leveraging high-frequency information. Moreover, diffusion-based approaches lack sufficient conditioning to fully utilize Panchromatic (PAN) images and low-resolution multispectral (LRMS) inputs effectively. To address these challenges, we propose an uncertainty-aware knowledge distillation diffusion framework with details enhancement for PAN-sharpening, called U-Know-DiffPAN. The U-Know-DiffPAN incorporates uncertainty-aware knowledge distillation for effective transfer of feature details from our teacher model to a student one. The teacher model in our U-Know-DiffPAN captures frequency details through freqeuncy selective attention, facilitating accurate reverse process learning. By conditioning the encoder on compact vector representations of PAN and LRMS and the decoder on Wavelet transforms, we enable rich frequency utilization. So, the high-capacity teacher model distills frequency-rich features into a lightweight student model aided by an uncertainty map. From this, the teacher model can guide the student model to focus on difficult image regions for PAN-sharpening via the usage of the uncertainty map. Extensive experiments on diverse datasets demonstrate the robustness and superior performance of our U-Know-DiffPAN over very recent state-of-the-art PAN-sharpening methods.
Augmented Reality (AR) is a major immersive media technology that enriches our perception of reality by overlaying digital content (the foreground) onto physical environments (the background). It has far-reaching applications, from entertainment and gaming to education, healthcare, and industrial training. Nevertheless, challenges such as visual confusion and classical distortions can result in user discomfort when using the technology. Evaluating AR quality of experience becomes essential to measure user satisfaction and engagement, facilitating the refinement necessary for creating immersive and robust experiences. Though, the scarcity of data and the distinctive characteristics of AR technology render the development of effective quality assessment metrics challenging. This paper presents a deep learning-based objective metric designed specifically for assessing image quality for AR scenarios. The approach entails four key steps, (1) fine-tuning a self-supervised pre-trained vision transformer to extract prominent features from reference images and distilling this knowledge to improve representations of distorted images, (2) quantifying distortions by computing shift representations, (3) employing cross-attention-based decoders to capture perceptual quality features, and (4) integrating regularization techniques and label smoothing to address the overfitting problem. To validate the proposed approach, we conduct extensive experiments on the ARIQA dataset. The results showcase the superior performance of our proposed approach across all model variants, namely TransformAR, TransformAR-KD, and TransformAR-KD+ in comparison to existing state-of-the-art methods.
Continuous Generalized Category Discovery (C-GCD) aims to continually discover novel classes from unlabelled image sets while maintaining performance on old classes. In this paper, we propose a novel learning framework, dubbed Neighborhood Commonality-aware Evolution Network (NCENet) that conquers this task from the perspective of representation learning. Concretely, to learn discriminative representations for novel classes, a Neighborhood Commonality-aware Representation Learning (NCRL) is designed, which exploits local commonalities derived neighborhoods to guide the learning of representational differences between instances of different classes. To maintain the representation ability for old classes, a Bi-level Contrastive Knowledge Distillation (BCKD) module is designed, which leverages contrastive learning to perceive the learning and learned knowledge and conducts knowledge distillation. Extensive experiments conducted on CIFAR10, CIFAR100, and Tiny-ImageNet demonstrate the superior performance of NCENet compared to the previous state-of-the-art method. Particularly, in the last incremental learning session on CIFAR100, the clustering accuracy of NCENet outperforms the second-best method by a margin of 3.09\% on old classes and by a margin of 6.32\% on new classes. Our code will be publicly available at \href{https://github.com/xjtuYW/NCENet.git}{https://github.com/xjtuYW/NCENet.git}. \end{abstract}
Federated Learning (FL) is an innovative distributed machine learning paradigm that enables neural network training across devices without centralizing data. While this addresses issues of information sharing and data privacy, challenges arise from data heterogeneity across clients and increasing network scale, leading to impacts on model performance and training efficiency. Previous research shows that in IID environments, the parameter structure of the model is expected to adhere to certain specific consistency principles. Thus, identifying and regularizing these consistencies can mitigate issues from heterogeneous data. We found that both soft labels derived from knowledge distillation and the classifier head parameter matrix, when multiplied by their own transpose, capture the intrinsic relationships between data classes. These shared relationships suggest inherent consistency. Therefore, the work in this paper identifies the consistency between the two and leverages it to regulate training, underpinning our proposed FedDW framework. Experimental results show FedDW outperforms 10 state-of-the-art FL methods, improving accuracy by an average of 3% in highly heterogeneous settings. Additionally, we provide a theoretical proof that FedDW offers higher efficiency, with the additional computational load from backpropagation being negligible. The code is available at https://github.com/liuvvvvv1/FedDW.
Wireless sensing has made significant progress in tasks ranging from action recognition, vital sign estimation, pose estimation, etc. After over a decade of work, wireless sensing currently stands at the tipping point transitioning from proof-of-concept systems to the large-scale deployment. We envision a future service scenario where wireless sensing service providers distribute sensing models to users. During usage, users might request new sensing capabilities. For example, if someone is away from home on a business trip or vacation for an extended period, they may want a new sensing capability that can detect falls in elderly parents or grandparents and promptly alert them. In this paper, we propose CCS (continuous customized service), enabling model updates on users' local computing resources without data transmission to the service providers. To address the issue of catastrophic forgetting in model updates where updating model parameters to implement new capabilities leads to the loss of existing capabilities we design knowledge distillation and weight alignment modules. These modules enable the sensing model to acquire new capabilities while retaining the existing ones. We conducted extensive experiments on the large-scale XRF55 dataset across Wi-Fi, millimeter-wave radar, and RFID modalities to simulate scenarios where four users sequentially introduced new customized demands. The results affirm that CCS excels in continuous model services across all the above wireless modalities, significantly outperforming existing approaches like OneFi.
One-shot Federated learning (FL) is a powerful technology facilitating collaborative training of machine learning models in a single round of communication. While its superiority lies in communication efficiency and privacy preservation compared to iterative FL, one-shot FL often compromises model performance. Prior research has primarily focused on employing data-free knowledge distillation to optimize data generators and ensemble models for better aggregating local knowledge into the server model. However, these methods typically struggle with data heterogeneity, where inconsistent local data distributions can cause teachers to provide misleading knowledge. Additionally, they may encounter scalability issues with complex datasets due to inherent two-step information loss: first, during local training (from data to model), and second, when transferring knowledge to the server model (from model to inversed data). In this paper, we propose FedSD2C, a novel and practical one-shot FL framework designed to address these challenges. FedSD2C introduces a distiller to synthesize informative distillates directly from local data to reduce information loss and proposes sharing synthetic distillates instead of inconsistent local models to tackle data heterogeneity. Our empirical results demonstrate that FedSD2C consistently outperforms other one-shot FL methods with more complex and real datasets, achieving up to 2.6 the performance of the best baseline. Code: https://github.com/Carkham/FedSD2C
With the rapid scaling of neural networks, data storage and communication demands have intensified. Dataset distillation has emerged as a promising solution, condensing information from extensive datasets into a compact set of synthetic samples by solving a bilevel optimization problem. However, current methods face challenges in computational efficiency, particularly with high-resolution data and complex architectures. Recently, knowledge-distillation-based dataset condensation approaches have made this process more computationally feasible. Yet, with the recent developments of generative foundation models, there is now an opportunity to achieve even greater compression, enhance the quality of distilled data, and introduce valuable diversity into the data representation. In this work, we propose a two-stage solution. First, we compress the dataset by selecting only the most informative patches to form a coreset. Next, we leverage a generative foundation model to dynamically expand this compressed set in real-time, enhancing the resolution of these patches and introducing controlled variability to the coreset. Our extensive experiments demonstrate the robustness and efficiency of our approach across a range of dataset distillation benchmarks. We demonstrate a significant improvement of over 10% compared to the state-of-the-art on several large-scale dataset distillation benchmarks. The code will be released soon.
Large Language Models (LLMs) based on transformers achieve cutting-edge results on a variety of applications. However, their enormous size and processing requirements make deployment on devices with constrained resources extremely difficult. Among various efficiency considerations, model binarization and Early Exit (EE) are common effective solutions. However, binarization may lead to performance loss due to reduced precision affecting gradient estimation and parameter updates. Besides, the present early-exit mechanisms are still in the nascent stages of research. To ameliorate these issues, we propose Binarized Early Exit Transformer (BEExformer), the first-ever selective learning transformer architecture to combine early exit with binarization for textual inference. It improves the binarization process through a differentiable second-order approximation to the impulse function. This enables gradient computation concerning both the sign as well as the magnitude of the weights. In contrast to absolute threshold-based EE, the proposed EE mechanism hinges on fractional reduction in entropy among intermediate transformer blocks with soft-routing loss estimation. While binarization results in 18.44 times reduction in model size, early exit reduces the FLOPs during inference by 54.85% and even improves accuracy by 5.98% through resolving the "overthinking" problem inherent in deep networks. Moreover, the proposed BEExformer simplifies training by not requiring knowledge distillation from a full-precision LLM. Extensive evaluation on the GLUE dataset and comparison with the SOTA works showcase its pareto-optimal performance-efficiency trade-off.
Expanding the existing sensing systems to provide high-quality deep learning models for more domains, such as new users or environments, is challenged by the limited labeled data and the data and device heterogeneities. While knowledge distillation methods could overcome label scarcity and device heterogeneity, they assume the teachers are fully reliable and overlook the data heterogeneity, which prevents the direct adoption of existing models. To address this problem, this paper proposes an efficient knowledge transfer framework, HaKT, to expand sensing systems. It first selects multiple high-quality models from the system at a low cost and then fuses their knowledge by assigning sample-wise weights to their predictions. Later, the fused knowledge is selectively injected into the customized models for new domains based on the knowledge quality. Extensive experiments on different tasks, modalities, and settings show that HaKT outperforms stat-of-the-art baselines by at most 16.5% accuracy and saves up to 39% communication traffic.
Semantic correspondence, the task of determining relationships between different parts of images, underpins various applications including 3D reconstruction, image-to-image translation, object tracking, and visual place recognition. Recent studies have begun to explore representations learned in large generative image models for semantic correspondence, demonstrating promising results. Building on this progress, current state-of-the-art methods rely on combining multiple large models, resulting in high computational demands and reduced efficiency. In this work, we address this challenge by proposing a more computationally efficient approach. We propose a novel knowledge distillation technique to overcome the problem of reduced efficiency. We show how to use two large vision foundation models and distill the capabilities of these complementary models into one smaller model that maintains high accuracy at reduced computational cost. Furthermore, we demonstrate that by incorporating 3D data, we are able to further improve performance, without the need for human-annotated correspondences. Overall, our empirical results demonstrate that our distilled model with 3D data augmentation achieves performance superior to current state-of-the-art methods while significantly reducing computational load and enhancing practicality for real-world applications, such as semantic video correspondence. Our code and weights are publicly available on our project page.
Recently, CLIP has become an important model for aligning images and text in multi-modal contexts. However, researchers have identified limitations in the ability of CLIP's text and image encoders to extract detailed knowledge from pairs of captions and images. In response, this paper presents Knowledge-CLIP, an innovative approach designed to improve CLIP's performance by integrating a new knowledge distillation (KD) method based on Llama 2. Our approach focuses on three key objectives: Text Embedding Distillation, Concept Learning, and Contrastive Learning. First, Text Embedding Distillation involves training the Knowledge-CLIP text encoder to mirror the teacher model, Llama 2. Next, Concept Learning assigns a soft concept label to each caption-image pair by employing offline K-means clustering on text data from Llama 2, enabling Knowledge-CLIP to learn from these soft concept labels. Lastly, Contrastive Learning aligns the text and image embeddings. Our experimental findings show that the proposed model improves the performance of both text and image encoders.
In this paper, we introduce QABISAR, a novel framework for statutory article retrieval, to overcome the semantic mismatch problem when modeling each query-article pair in isolation, making it hard to learn representation that can effectively capture multi-faceted information. QABISAR leverages bipartite interactions between queries and articles to capture diverse aspects inherent in them. Further, we employ knowledge distillation to transfer enriched query representations from the graph network into the query bi-encoder, to capture the rich semantics present in the graph representations, despite absence of graph-based supervision for unseen queries during inference. Our experiments on a real-world expert-annotated dataset demonstrate its effectiveness.
Graph Neural Networks (GNNs) have demonstrated strong performance in graph representation learning across various real-world applications. However, they often produce biased predictions caused by sensitive attributes, such as religion or gender, an issue that has been largely overlooked in existing methods. Recently, numerous studies have focused on reducing biases in GNNs. However, these approaches often rely on training with partial data (e.g., using either node features or graph structure alone), which can enhance fairness but frequently compromises model utility due to the limited utilization of available graph information. To address this tradeoff, we propose an effective strategy to balance fairness and utility in knowledge distillation. Specifically, we introduce FairDTD, a novel Fair representation learning framework built on Dual-Teacher Distillation, leveraging a causal graph model to guide and optimize the design of the distillation process. Specifically, FairDTD employs two fairness-oriented teacher models: a feature teacher and a structure teacher, to facilitate dual distillation, with the student model learning fairness knowledge from the teachers while also leveraging full data to mitigate utility loss. To enhance information transfer, we incorporate graph-level distillation to provide an indirect supplement of graph information during training, as well as a node-specific temperature module to improve the comprehensive transfer of fair knowledge. Experiments on diverse benchmark datasets demonstrate that FairDTD achieves optimal fairness while preserving high model utility, showcasing its effectiveness in fair representation learning for GNNs.
In 2023, 58.0% of the African population experienced moderate to severe food insecurity, with 21.6% facing severe food insecurity. Land-use and land-cover maps provide crucial insights for addressing food insecurity by improving agricultural efforts, including mapping and monitoring crop types and estimating yield. The development of global land-cover maps has been facilitated by the increasing availability of earth observation data and advancements in geospatial machine learning. However, these global maps exhibit lower accuracy and inconsistencies in Africa, partly due to the lack of representative training data. To address this issue, we propose a data-centric framework with a teacher-student model setup, which uses diverse data sources of satellite images and label examples to produce local land-cover maps. Our method trains a high-resolution teacher model on images with a resolution of 0.331 m/pixel and a low-resolution student model on publicly available images with a resolution of 10 m/pixel. The student model also utilizes the teacher model's output as its weak label examples through knowledge transfer. We evaluated our framework using Murang'a county in Kenya, renowned for its agricultural productivity, as a use case. Our local models achieved higher quality maps, with improvements of 0.14 in the F1 score and 0.21 in Intersection-over-Union, compared to the best global model. Our evaluation also revealed inconsistencies in existing global maps, with a maximum agreement rate of 0.30 among themselves. Our work provides valuable guidance to decision-makers for driving informed decisions to enhance food security.
Large Foundation Models like Dust3r can produce high quality outputs such as pointmaps, camera intrinsics, and depth estimation, given stereo-image pairs as input. However, the application of these outputs on tasks like Visual Localization requires a large amount of inference time and compute resources. To address these limitations, in this paper, we propose the use of a knowledge distillation pipeline, where we aim to build a student-teacher model with Dust3r as the teacher and explore multiple architectures of student models that are trained using the 3D reconstructed points output by Dust3r. Our goal is to build student models that can learn scene-specific representations and output 3D points with replicable performance such as Dust3r. The data set we used to train our models is 12Scenes. We test two main architectures of models: a CNN-based architecture and a Vision Transformer based architecture. For each architecture, we also compare the use of pre-trained models against models built from scratch. We qualitatively compare the reconstructed 3D points output by the student model against Dust3r's and discuss the various features learned by the student model. We also perform ablation studies on the models through hyperparameter tuning. Overall, we observe that the Vision Transformer presents the best performance visually and quantitatively.