Recent advances in text-to-image generation have primarily relied on extensive datasets and parameter-heavy architectures. These requirements severely limit accessibility for researchers and practitioners who lack substantial computational resources. In this paper, we introduce \model, an efficient training paradigm for image generation models that uses knowledge distillation (KD) and Direct Preference Optimization (DPO). Drawing inspiration from the success of data KD techniques widely adopted in Multi-Modal Large Language Models (MLLMs), LightGen distills knowledge from state-of-the-art (SOTA) text-to-image models into a compact Masked Autoregressive (MAR) architecture with only $0.7B$ parameters. Using a compact synthetic dataset of just $2M$ high-quality images generated from varied captions, we demonstrate that data diversity significantly outweighs data volume in determining model performance. This strategy dramatically reduces computational demands and reduces pre-training time from potentially thousands of GPU-days to merely 88 GPU-days. Furthermore, to address the inherent shortcomings of synthetic data, particularly poor high-frequency details and spatial inaccuracies, we integrate the DPO technique that refines image fidelity and positional accuracy. Comprehensive experiments confirm that LightGen achieves image generation quality comparable to SOTA models while significantly reducing computational resources and expanding accessibility for resource-constrained environments. Code is available at https://github.com/XianfengWu01/LightGen
Large-scale pre-trained models, such as Vision Foundation Models (VFMs), have demonstrated impressive performance across various downstream tasks by transferring generalized knowledge, especially when target data is limited. However, their high computational cost and the domain gap between natural and medical images limit their practical application in medical segmentation tasks. Motivated by this, we pose the following important question: "How can we effectively utilize the knowledge of large pre-trained VFMs to train a small, task-specific model for medical image segmentation when training data is limited?" To address this problem, we propose a novel and generalizable task-specific knowledge distillation framework. Our method fine-tunes the VFM on the target segmentation task to capture task-specific features before distilling the knowledge to smaller models, leveraging Low-Rank Adaptation (LoRA) to reduce the computational cost of fine-tuning. Additionally, we incorporate synthetic data generated by diffusion models to augment the transfer set, enhancing model performance in data-limited scenarios. Experimental results across five medical image datasets demonstrate that our method consistently outperforms task-agnostic knowledge distillation and self-supervised pretraining approaches like MoCo v3 and Masked Autoencoders (MAE). For example, on the KidneyUS dataset, our method achieved a 28% higher Dice score than task-agnostic KD using 80 labeled samples for fine-tuning. On the CHAOS dataset, it achieved an 11% improvement over MAE with 100 labeled samples. These results underscore the potential of task-specific knowledge distillation to train accurate, efficient models for medical image segmentation in data-constrained settings.
Online Knowledge Distillation (OKD) methods streamline the distillation training process into a single stage, eliminating the need for knowledge transfer from a pretrained teacher network to a more compact student network. This paper presents an innovative approach to leverage intermediate spatial representations. Our analysis of the intermediate features from both teacher and student models reveals two pivotal insights: (1) the similar features between students and teachers are predominantly focused on foreground objects. (2) teacher models emphasize foreground objects more than students. Building on these findings, we propose Asymmetric Decision-Making (ADM) to enhance feature consensus learning for student models while continuously promoting feature diversity in teacher models. Specifically, Consensus Learning for student models prioritizes spatial features with high consensus relative to teacher models. Conversely, Divergence Learning for teacher models highlights spatial features with lower similarity compared to student models, indicating superior performance by teacher models in these regions. Consequently, ADM facilitates the student models to catch up with the feature learning process of the teacher models. Extensive experiments demonstrate that ADM consistently surpasses existing OKD methods across various online knowledge distillation settings and also achieves superior results when applied to offline knowledge distillation, semantic segmentation and diffusion distillation tasks.
Dense visual prediction tasks, such as detection and segmentation, are crucial for time-critical applications (e.g., autonomous driving and video surveillance). While deep models achieve strong performance, their efficiency remains a challenge. Knowledge distillation (KD) is an effective model compression technique, but existing feature-based KD methods rely on static, teacher-driven feature selection, failing to adapt to the student's evolving learning state or leverage dynamic student-teacher interactions. To address these limitations, we propose Adaptive student-teacher Cooperative Attention Masking for Knowledge Distillation (ACAM-KD), which introduces two key components: (1) Student-Teacher Cross-Attention Feature Fusion (STCA-FF), which adaptively integrates features from both models for a more interactive distillation process, and (2) Adaptive Spatial-Channel Masking (ASCM), which dynamically generates importance masks to enhance both spatial and channel-wise feature selection. Unlike conventional KD methods, ACAM-KD adapts to the student's evolving needs throughout the entire distillation process. Extensive experiments on multiple benchmarks validate its effectiveness. For instance, on COCO2017, ACAM-KD improves object detection performance by up to 1.4 mAP over the state-of-the-art when distilling a ResNet-50 student from a ResNet-101 teacher. For semantic segmentation on Cityscapes, it boosts mIoU by 3.09 over the baseline with DeepLabV3-MobileNetV2 as the student model.
Segment anything model (SAM) has shown impressive general-purpose segmentation performance on natural images, but its performance on camouflaged object detection (COD) is unsatisfactory. In this paper, we propose SAM-COD that performs camouflaged object detection for RGB-D inputs. While keeping the SAM architecture intact, dual stream adapters are expanded on the image encoder to learn potential complementary information from RGB images and depth images, and fine-tune the mask decoder and its depth replica to perform dual-stream mask prediction. In practice, the dual stream adapters are embedded into the attention block of the image encoder in a parallel manner to facilitate the refinement and correction of the two types of image embeddings. To mitigate channel discrepancies arising from dual stream embeddings that do not directly interact with each other, we augment the association of dual stream embeddings using bidirectional knowledge distillation including a model distiller and a modal distiller. In addition, to predict the masks for RGB and depth attention maps, we hybridize the two types of image embeddings which are jointly learned with the prompt embeddings to update the initial prompt, and then feed them into the mask decoders to synchronize the consistency of image embeddings and prompt embeddings. Experimental results on four COD benchmarks show that our SAM-COD achieves excellent detection performance gains over SAM and achieves state-of-the-art results with a given fine-tuning paradigm.
Learnable Image Compression (LIC) has shown the potential to outperform standardized video codecs in RD efficiency, prompting the research for hardware-friendly implementations. Most existing LIC hardware implementations prioritize latency to RD-efficiency and through an extensive exploration of the hardware design space. We present a novel design paradigm where the burden of tuning the design for a specific hardware platform is shifted towards model dimensioning and without compromising on RD-efficiency. First, we design a framework for distilling a leaner student LIC model from a reference teacher: by tuning a single model hyperparameters, we can meet the constraints of different hardware platforms without a complex hardware design exploration. Second, we propose a hardware-friendly implementation of the Generalized Divisive Normalization - GDN activation that preserves RD efficiency even post parameter quantization. Third, we design a pipelined FPGA configuration which takes full advantage of available FPGA resources by leveraging parallel processing and optimizing resource allocation. Our experiments with a state of the art LIC model show that we outperform all existing FPGA implementations while performing very close to the original model.
Active learning aims to select optimal samples for labeling, minimizing annotation costs. This paper introduces a unified representation learning framework tailored for active learning with task awareness. It integrates diverse sources, comprising reconstruction, adversarial, self-supervised, knowledge-distillation, and classification losses into a unified VAE-based ADROIT approach. The proposed approach comprises three key components - a unified representation generator (VAE), a state discriminator, and a (proxy) task-learner or classifier. ADROIT learns a latent code using both labeled and unlabeled data, incorporating task-awareness by leveraging labeled data with the proxy classifier. Unlike previous approaches, the proxy classifier additionally employs a self-supervised loss on unlabeled data and utilizes knowledge distillation to align with the target task-learner. The state discriminator distinguishes between labeled and unlabeled data, facilitating the selection of informative unlabeled samples. The dynamic interaction between VAE and the state discriminator creates a competitive environment, with the VAE attempting to deceive the discriminator, while the state discriminator learns to differentiate between labeled and unlabeled inputs. Extensive evaluations on diverse datasets and ablation analysis affirm the effectiveness of the proposed model.
Quantum vision transformers (QViTs) build on vision transformers (ViTs) by replacing linear layers within the self-attention mechanism with parameterised quantum neural networks (QNNs), harnessing quantum mechanical properties to improve feature representation. This hybrid approach aims to achieve superior performance, with significantly reduced model complexity as a result of the enriched feature representation, requiring fewer parameters. This paper proposes a novel QViT model for biomedical image classification and investigates its performance against comparable ViTs across eight diverse datasets, encompassing various modalities and classification tasks. We assess models trained from scratch and those pre-trained using knowledge distillation (KD) from high-quality teacher models. Our findings demonstrate that QViTs outperform comparable ViTs with average ROC AUC (0.863 vs 0.846) and accuracy (0.710 vs 0.687) when trained from scratch, and even compete with state-of-the-art classical models in multiple tasks, whilst being significantly more efficient (89% reduction in GFLOPs and 99.99% in parameter number). Additionally, we find that QViTs and ViTs respond equally well to KD, with QViT pre-training performance scaling with model complexity. This is the first investigation into the efficacy of deploying QViTs with KD for computer-aided diagnosis. Our results highlight the enormous potential of quantum machine learning (QML) in biomedical image analysis.
Class-incremental learning (CIL) for time series data faces critical challenges in balancing stability against catastrophic forgetting and plasticity for new knowledge acquisition, particularly under real-world constraints where historical data access is restricted. While pre-trained models (PTMs) have shown promise in CIL for vision and NLP domains, their potential in time series class-incremental learning (TSCIL) remains underexplored due to the scarcity of large-scale time series pre-trained models. Prompted by the recent emergence of large-scale pre-trained models (PTMs) for time series data, we present the first exploration of PTM-based Time Series Class-Incremental Learning (TSCIL). Our approach leverages frozen PTM backbones coupled with incrementally tuning the shared adapter, preserving generalization capabilities while mitigating feature drift through knowledge distillation. Furthermore, we introduce a Feature Drift Compensation Network (DCN), designed with a novel two-stage training strategy to precisely model feature space transformations across incremental tasks. This allows for accurate projection of old class prototypes into the new feature space. By employing DCN-corrected prototypes, we effectively enhance the unified classifier retraining, mitigating model feature drift and alleviating catastrophic forgetting. Extensive experiments on five real-world datasets demonstrate state-of-the-art performance, with our method yielding final accuracy gains of 1.4%-6.1% across all datasets compared to existing PTM-based approaches. Our work establishes a new paradigm for TSCIL, providing insights into stability-plasticity optimization for continual learning systems.
Most current federated learning frameworks are modeled as static processes, ignoring the dynamic characteristics of the learning system. Under the limited communication budget of the central server, the flexible model architecture of a large number of clients participating in knowledge transfer requires a lower participation rate, active clients have uneven contributions, and the client scale seriously hinders the performance of FL. We consider a more general and practical federation scenario and propose a system heterogeneous federation method based on data-free knowledge distillation and two-way contrast (HFedCKD). We apply the Inverse Probability Weighted Distillation (IPWD) strategy to the data-free knowledge transfer framework. The generator completes the data features of the nonparticipating clients. IPWD implements a dynamic evaluation of the prediction contribution of each client under different data distributions. Based on the antibiased weighting of its prediction loss, the weight distribution of each client is effectively adjusted to fairly integrate the knowledge of participating clients. At the same time, the local model is split into a feature extractor and a classifier. Through differential contrast learning, the feature extractor is aligned with the global model in the feature space, while the classifier maintains personalized decision-making capabilities. HFedCKD effectively alleviates the knowledge offset caused by a low participation rate under data-free knowledge distillation and improves the performance and stability of the model. We conduct extensive experiments on image and IoT datasets to comprehensively evaluate and verify the generalization and robustness of the proposed HFedCKD framework.
Accurate origin-destination (OD) flow prediction is of great importance to developing cities, as it can contribute to optimize urban structures and layouts. However, with the common issues of missing regional features and lacking OD flow data, it is quite daunting to predict OD flow in developing cities. To address this challenge, we propose a novel Causality-Enhanced OD Flow Prediction (CE-OFP), a unified framework that aims to transfer urban knowledge between cities and achieve accuracy improvements in OD flow predictions across data-scarce cities. In specific, we propose a novel reinforcement learning model to discover universal causalities among urban features in data-rich cities and build corresponding causal graphs. Then, we further build Causality-Enhanced Variational Auto-Encoder (CE-VAE) to incorporate causal graphs for effective feature reconstruction in data-scarce cities. Finally, with the reconstructed features, we devise a knowledge distillation method with a graph attention network to migrate the OD prediction model from data-rich cities to data-scare cities. Extensive experiments on two pairs of real-world datasets validate that the proposed CE-OFP remarkably outperforms state-of-the-art baselines, which can reduce the RMSE of OD flow prediction for data-scarce cities by up to 11%.
Three-dimensional biological microscopy has significantly advanced our understanding of complex biological structures. However, limitations due to microscopy techniques, sample properties or phototoxicity often result in poor z-resolution, hindering accurate cellular measurements. Here, we introduce ZAugNet, a fast, accurate, and self-supervised deep learning method for enhancing z-resolution in biological images. By performing nonlinear interpolation between consecutive slices, ZAugNet effectively doubles resolution with each iteration. Compared on several microscopy modalities and biological objects, it outperforms competing methods on most metrics. Our method leverages a generative adversarial network (GAN) architecture combined with knowledge distillation to maximize prediction speed without compromising accuracy. We also developed ZAugNet+, an extended version enabling continuous interpolation at arbitrary distances, making it particularly useful for datasets with nonuniform slice spacing. Both ZAugNet and ZAugNet+ provide high-performance, scalable z-slice augmentation solutions for large-scale 3D imaging. They are available as open-source frameworks in PyTorch, with an intuitive Colab notebook interface for easy access by the scientific community.
Accurate motion forecasting is crucial for safe autonomous driving (AD). This study proposes CoT-Drive, a novel approach that enhances motion forecasting by leveraging large language models (LLMs) and a chain-of-thought (CoT) prompting method. We introduce a teacher-student knowledge distillation strategy to effectively transfer LLMs' advanced scene understanding capabilities to lightweight language models (LMs), ensuring that CoT-Drive operates in real-time on edge devices while maintaining comprehensive scene understanding and generalization capabilities. By leveraging CoT prompting techniques for LLMs without additional training, CoT-Drive generates semantic annotations that significantly improve the understanding of complex traffic environments, thereby boosting the accuracy and robustness of predictions. Additionally, we present two new scene description datasets, Highway-Text and Urban-Text, designed for fine-tuning lightweight LMs to generate context-specific semantic annotations. Comprehensive evaluations of five real-world datasets demonstrate that CoT-Drive outperforms existing models, highlighting its effectiveness and efficiency in handling complex traffic scenarios. Overall, this study is the first to consider the practical application of LLMs in this field. It pioneers the training and use of a lightweight LLM surrogate for motion forecasting, setting a new benchmark and showcasing the potential of integrating LLMs into AD systems.
A non-invasive brain-computer interface (BCI) enables direct interaction between the user and external devices, typically via electroencephalogram (EEG) signals. However, decoding EEG signals across different headsets remains a significant challenge due to differences in the number and locations of the electrodes. To address this challenge, we propose a spatial distillation based distribution alignment (SDDA) approach for heterogeneous cross-headset transfer in non-invasive BCIs. SDDA uses first spatial distillation to make use of the full set of electrodes, and then input/feature/output space distribution alignments to cope with the significant differences between the source and target domains. To our knowledge, this is the first work to use knowledge distillation in cross-headset transfers. Extensive experiments on six EEG datasets from two BCI paradigms demonstrated that SDDA achieved superior performance in both offline unsupervised domain adaptation and online supervised domain adaptation scenarios, consistently outperforming 10 classical and state-of-the-art transfer learning algorithms.
Transformer-based and CNN-based methods demonstrate strong performance in long-term time series forecasting. However, their high computational and storage requirements can hinder large-scale deployment. To address this limitation, we propose integrating lightweight MLP with advanced architectures using knowledge distillation (KD). Our preliminary study reveals different models can capture complementary patterns, particularly multi-scale and multi-period patterns in the temporal and frequency domains. Based on this observation, we introduce TimeDistill, a cross-architecture KD framework that transfers these patterns from teacher models (e.g., Transformers, CNNs) to MLP. Additionally, we provide a theoretical analysis, demonstrating that our KD approach can be interpreted as a specialized form of mixup data augmentation. TimeDistill improves MLP performance by up to 18.6%, surpassing teacher models on eight datasets. It also achieves up to 7X faster inference and requires 130X fewer parameters. Furthermore, we conduct extensive evaluations to highlight the versatility and effectiveness of TimeDistill.
Spiking Neural Networks (SNNs), inspired by the human brain, offer significant computational efficiency through discrete spike-based information transfer. Despite their potential to reduce inference energy consumption, a performance gap persists between SNNs and Artificial Neural Networks (ANNs), primarily due to current training methods and inherent model limitations. While recent research has aimed to enhance SNN learning by employing knowledge distillation (KD) from ANN teacher networks, traditional distillation techniques often overlook the distinctive spatiotemporal properties of SNNs, thus failing to fully leverage their advantages. To overcome these challenge, we propose a novel logit distillation method characterized by temporal separation and entropy regularization. This approach improves existing SNN distillation techniques by performing distillation learning on logits across different time steps, rather than merely on aggregated output features. Furthermore, the integration of entropy regularization stabilizes model optimization and further boosts the performance. Extensive experimental results indicate that our method surpasses prior SNN distillation strategies, whether based on logit distillation, feature distillation, or a combination of both. The code will be available on GitHub.
Rapid bone scintigraphy is crucial for diagnosing skeletal disorders and detecting tumor metastases in children, as it shortens scan duration and reduces discomfort. However, accelerated acquisition often degrades image quality, impairing the visibility of fine anatomical details and potentially compromising diagnosis. To overcome this limitation, we introduce the first application of SAM-based semantic priors for medical image restoration, utilizing the Segment Anything Model (SAM) to enhance pediatric rapid bone scintigraphy. Our approach employs two cascaded networks, $f^{IR1}$ and $f^{IR2}$, supported by three specialized modules: a Semantic Prior Integration (SPI) module, a Semantic Knowledge Distillation (SKD) module, and a Semantic Consistency Module (SCM). The SPI and SKD modules inject domain-specific semantic cues from a fine-tuned SAM, while the SCM preserves coherent semantic feature representations across both cascaded stages. Moreover, we present RBS, a novel Rapid Bone Scintigraphy dataset comprising paired standard (20 cm/min) and rapid (40 cm/min) scans from 137 pediatric patients aged 0.5 - 16 years, making it the first dataset tailored for pediatric rapid bone scintigraphy restoration. Extensive experiments on both a public endoscopic dataset and our RBS dataset demonstrate that our method consistently surpasses existing techniques in PSNR, SSIM, FID, and LPIPS metrics.
Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100 and ImageNet datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method over a range of models, architectures, and set-ups. For instance, VRM for the first time hits 74.0% accuracy for ResNet50-to-MobileNetV2 distillation on ImageNet, and improves DeiT-T by 14.44% on CIFAR-100 with a ResNet56 teacher. Thorough analyses are also conducted to gauge the soundness, properties, and complexity of our designs. Code and models will be released.
Autoformalization, the automatic translation of mathematical content from natural language into machine-verifiable formal languages, has seen significant progress driven by advances in large language models (LLMs). Nonetheless, a primary barrier to further improvements is the limited availability of parallel corpora that map informal mathematical text to its formal counterpart. To address this limitation, we propose ATLAS (Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data), a novel data generation framework designed to produce large-scale, high-quality parallel corpora of theorem statements. Distinct from prior approaches, ATLAS begins with a concept repository, accelerates the improvement of student model through expert iteration combined with knowledge distillation, and introduces two novel augmentation strategies that exploit the structural characteristics of formal languages. With the proposed ATLAS running for 10 iterations, we construct an undergraduate-level dataset comprising 117k theorem statements and develop ATLAS Translator, which demonstrates statistically significant improvements over both the HERALD Translator and the Kimina-Autoformalizer across all benchmarks ($p<0.05$, two-sided t-test), achieving a new state of the art. The datasets, model, and code will be released to the public soon.
Deep neural networks (DNNs) have remarkably succeeded in various image processing tasks. However, their large size and computational complexity present significant challenges for deploying them in resource-constrained environments. This paper presents an innovative approach for integrating Mamba Architecture within a Progressive Knowledge Distillation (PKD) process to address the challenge of reducing model complexity while maintaining accuracy in image classification tasks. The proposed framework distills a large teacher model into progressively smaller student models, designed using Mamba blocks. Each student model is trained using Selective-State-Space Models (S-SSM) within the Mamba blocks, focusing on important input aspects while reducing computational complexity. The work's preliminary experiments use MNIST and CIFAR-10 as datasets to demonstrate the effectiveness of this approach. For MNIST, the teacher model achieves 98% accuracy. A set of seven student models as a group retained 63% of the teacher's FLOPs, approximating the teacher's performance with 98% accuracy. The weak student used only 1% of the teacher's FLOPs and maintained 72% accuracy. Similarly, for CIFAR-10, the students achieved 1% less accuracy compared to the teacher, with the small student retaining 5% of the teacher's FLOPs to achieve 50% accuracy. These results confirm the flexibility and scalability of Mamba Architecture, which can be integrated into PKD, succeeding in the process of finding students as weak learners. The framework provides a solution for deploying complex neural networks in real-time applications with a reduction in computational cost.