Diffusion models generate high-quality images through progressive denoising but are computationally intensive due to large model sizes and repeated sampling. Knowledge distillation, which transfers knowledge from a complex teacher to a simpler student model, has been widely studied in recognition tasks, particularly for transferring concepts unseen during student training. However, its application to diffusion models remains underexplored, especially in enabling student models to generate concepts not covered by the training images. In this work, we propose Random Conditioning, a novel approach that pairs noised images with randomly selected text conditions to enable efficient, image-free knowledge distillation. By leveraging this technique, we show that the student can generate concepts unseen in the training images. When applied to conditional diffusion model distillation, our method allows the student to explore the condition space without generating condition-specific images, resulting in notable improvements in both generation quality and efficiency. This promotes resource-efficient deployment of generative diffusion models, broadening their accessibility for both research and real-world applications. Code, models, and datasets are available at https://dohyun-as.github.io/Random-Conditioning .
Marine Saliency Segmentation (MSS) plays a pivotal role in various vision-based marine exploration tasks. However, existing marine segmentation techniques face the dilemma of object mislocalization and imprecise boundaries due to the complex underwater environment. Meanwhile, despite the impressive performance of diffusion models in visual segmentation, there remains potential to further leverage contextual semantics to enhance feature learning of region-level salient objects, thereby improving segmentation outcomes. Building on this insight, we propose DiffMSS, a novel marine saliency segmenter based on the diffusion model, which utilizes semantic knowledge distillation to guide the segmentation of marine salient objects. Specifically, we design a region-word similarity matching mechanism to identify salient terms at the word level from the text descriptions. These high-level semantic features guide the conditional feature learning network in generating salient and accurate diffusion conditions with semantic knowledge distillation. To further refine the segmentation of fine-grained structures in unique marine organisms, we develop the dedicated consensus deterministic sampling to suppress overconfident missegmentations. Comprehensive experiments demonstrate the superior performance of DiffMSS over state-of-the-art methods in both quantitative and qualitative evaluations.
The deployment of foundation models for medical imaging has demonstrated considerable success. However, their training overheads associated with downstream tasks remain substantial due to the size of the image encoders employed, and the inference complexity is also significantly high. Although lightweight variants have been obtained for these foundation models, their performance is constrained by their limited model capacity and suboptimal training strategies. In order to achieve an improved tradeoff between complexity and performance, we propose a new framework to improve the performance of low complexity models via knowledge distillation from multiple large medical foundation models (e.g., MedSAM, RAD-DINO, MedCLIP), each specializing in different vision tasks, with the goal to effectively bridge the performance gap for medical image segmentation tasks. The agglomerated model demonstrates superior generalization across 12 segmentation tasks, whereas specialized models require explicit training for each task. Our approach achieved an average performance gain of 2\% in Dice coefficient compared to simple distillation.
Speech separation (SS) seeks to disentangle a multi-talker speech mixture into single-talker speech streams. Although SS can be generally achieved using offline methods, such a processing paradigm is not suitable for real-time streaming applications. Causal separation models, which rely only on past and present information, offer a promising solution for real-time streaming. However, these models typically suffer from notable performance degradation due to the absence of future context. In this paper, we introduce a novel frontend that is designed to mitigate the mismatch between training and run-time inference by implicitly incorporating future information into causal models through predictive patterns. The pretrained frontend employs a transformer decoder network with a causal convolutional encoder as the backbone and is pretrained in a self-supervised manner with two innovative pretext tasks: autoregressive hybrid prediction and contextual knowledge distillation. These tasks enable the model to capture predictive patterns directly from mixtures in a self-supervised manner. The pretrained frontend subsequently serves as a feature extractor to generate high-quality predictive patterns. Comprehensive evaluations on synthetic and real-world datasets validated the effectiveness of the proposed pretrained frontend.
Medical imaging, particularly X-ray analysis, often involves detecting multiple conditions simultaneously within a single scan, making multi-label classification crucial for real-world clinical applications. We present the Medical X-ray Attention (MXA) block, a novel attention mechanism tailored specifically to address the unique challenges of X-ray abnormality detection. The MXA block enhances traditional Multi-Head Self Attention (MHSA) by integrating a specialized module that efficiently captures both detailed local information and broader global context. To the best of our knowledge, this is the first work to propose a task-specific attention mechanism for diagnosing chest X-rays, as well as to attempt multi-label classification using an Efficient Vision Transformer (EfficientViT). By embedding the MXA block within the EfficientViT architecture and employing knowledge distillation, our proposed model significantly improves performance on the CheXpert dataset, a widely used benchmark for multi-label chest X-ray abnormality detection. Our approach achieves an area under the curve (AUC) of 0.85, an absolute improvement of 0.19 compared to our baseline model's AUC of 0.66, corresponding to a substantial approximate 233% relative improvement over random guessing (AUC = 0.5).
With Neural Radiance Fields (NeRFs) arising as a powerful 3D representation, research has investigated its various downstream tasks, including inpainting NeRFs with 2D images. Despite successful efforts addressing the view consistency and geometry quality, prior methods yet suffer from occlusion in NeRF inpainting tasks, where 2D prior is severely limited in forming a faithful reconstruction of the scene to inpaint. To address this, we propose a novel approach that enables cross-view information sharing during knowledge distillation from a diffusion model, effectively propagating occluded information across limited views. Additionally, to align the distillation direction across multiple sampled views, we apply a grid-based denoising strategy and incorporate additional rendered views to enhance cross-view consistency. To assess our approach's capability of handling occlusion cases, we construct a dataset consisting of challenging scenes with severe occlusion, in addition to existing datasets. Compared with baseline methods, our method demonstrates better performance in cross-view consistency and faithfulness in reconstruction, while preserving high rendering quality and fidelity.
Specialized reasoning language models (RLMs) have demonstrated that scaling test-time computation through detailed reasoning traces significantly enhances performance. Although these traces effectively facilitate knowledge distillation into smaller, instruction-tuned models, the precise nature of transferred reasoning remains unclear. In this study, we investigate to what extent distilled models internalize replicated stylistic patterns during reasoning. To this end, we systematically analyze reasoning traces, identifying structural and lexical patterns that characterize successful reasoning. We then introduce two new datasets -- a dataset of emergent reasoning traces and a synthetic dataset explicitly constructed to replicate these stylistic patterns -- to precisely examine their influence on distilled models' reasoning capabilities. We find that models trained on the synthetic traces achieve comparable performance, indicating that distilled reasoning abilities rely significantly on surface-level patterns. Surprisingly, we observe an increase in performance even when the synthetic traces are altered to lead to the wrong answer. Our findings highlight how stylistic patterns can be leveraged to efficiently enhance LM reasoning across diverse model families.
Knowledge Distillation (KD) seeks to transfer the knowledge of a teacher, towards a student neural net. This process is often done by matching the networks' predictions (i.e., their output), but, recently several works have proposed to match the distributions of neural nets' activations (i.e., their features), a process known as \emph{distribution matching}. In this paper, we propose an unifying framework, Knowledge Distillation through Distribution Matching (KD$^{2}$M), which formalizes this strategy. Our contributions are threefold. We i) provide an overview of distribution metrics used in distribution matching, ii) benchmark on computer vision datasets, and iii) derive new theoretical results for KD.
Data-free Knowledge Distillation (DFKD) is a method that constructs pseudo-samples using a generator without real data, and transfers knowledge from a teacher model to a student by enforcing the student to overcome dimensional differences and learn to mimic the teacher's outputs on these pseudo-samples. In recent years, various studies in the vision domain have made notable advancements in this area. However, the varying topological structures and non-grid nature of graph data render the methods from the vision domain ineffective. Building upon prior research into differentiable methods for graph neural networks, we propose a fast and high-quality data-free knowledge distillation approach in this paper. Without compromising distillation quality, the proposed graph-free KD method (ACGKD) significantly reduces the spatial complexity of pseudo-graphs by leveraging the Binary Concrete distribution to model the graph structure and introducing a spatial complexity tuning parameter. This approach enables efficient gradient computation for the graph structure, thereby accelerating the overall distillation process. Additionally, ACGKD eliminates the dimensional ambiguity between the student and teacher models by increasing the student's dimensions and reusing the teacher's classifier. Moreover, it equips graph knowledge distillation with a CL-based strategy to ensure the student learns graph structures progressively. Extensive experiments demonstrate that ACGKD achieves state-of-the-art performance in distilling knowledge from GNNs without training data.
The Tsetlin Machine (TM) is a propositional logic based model that uses conjunctive clauses to learn patterns from data. As with typical neural networks, the performance of a Tsetlin Machine is largely dependent on its parameter count, with a larger number of parameters producing higher accuracy but slower execution. Knowledge distillation in neural networks transfers information from an already-trained teacher model to a smaller student model to increase accuracy in the student without increasing execution time. We propose a novel approach to implementing knowledge distillation in Tsetlin Machines by utilizing the probability distributions of each output sample in the teacher to provide additional context to the student. Additionally, we propose a novel clause-transfer algorithm that weighs the importance of each clause in the teacher and initializes the student with only the most essential data. We find that our algorithm can significantly improve performance in the student model without negatively impacting latency in the tested domains of image recognition and text classification.
Real-time eyeblink detection in the wild is a recently emerged challenging task that suffers from dramatic variations in face attribute, pose, illumination, camera view and distance, etc. One key issue is to well characterize eyelid’s intrinsic motion (i.e., approaching and departure between upper and lower eyelid) robustly, under unconstrained conditions. Towards this, a novel eyelid’s intrinsic motion-aware feature learning approach is proposed. Our proposition lies in 3 folds. First, the feature extractor is led to focus on informative eye region adaptively via introducing visual attention in a coarse-to-fine way, to guarantee robustness and fine-grained descriptive ability jointly. Then, 2 constraints are proposed to make feature learning be aware of eyelid’s intrinsic motion. Particularly, one concerns the fact that the inter-frame feature divergence within eyeblink processes should be greater than non-eyeblink ones to better reveal eyelid’s intrinsic motion. The other constraint minimizes the inter-frame feature divergence of non-eyeblink samples, to suppress motion clues due to head or camera movement, illumination change, etc. Meanwhile, concerning the high ambiguity between eyeblink and non-eyeblink samples, soft sample labels are acquired via self-knowledge distillation to conduct feature learning with finer supervision than the hard ones. The experiments verify that, our proposition is significantly superior to the state-of-the-art ones, and with real-time running efficiency. It is also of strong generalization capacity towards constrained conditions.
Low-Resource Languages (LRLs) present significant challenges in natural language processing due to their limited linguistic resources and underrepresentation in standard datasets. While recent advancements in Large Language Models (LLMs) and Neural Machine Translation (NMT) have substantially improved translation capabilities for high-resource languages, performance disparities persist for LRLs, particularly impacting privacy-sensitive and resource-constrained scenarios. This paper systematically evaluates the limitations of current LLMs across 200 languages using benchmarks such as FLORES-200. We also explore alternative data sources, including news articles and bilingual dictionaries, and demonstrate how knowledge distillation from large pre-trained models can significantly improve smaller LRL translations. Additionally, we investigate various fine-tuning strategies, revealing that incremental enhancements markedly reduce performance gaps on smaller LLMs.
Crossmodal knowledge distillation (KD) aims to enhance a unimodal student using a multimodal teacher model. In particular, when the teacher's modalities include the student's, additional complementary information can be exploited to improve knowledge transfer. In supervised image classification, image datasets typically include class labels that represent high-level concepts, suggesting a natural avenue to incorporate textual cues for crossmodal KD. However, these labels rarely capture the deeper semantic structures in real-world visuals and can lead to label leakage if used directly as inputs, ultimately limiting KD performance. To address these issues, we propose a multi-teacher crossmodal KD framework that integrates CLIP image embeddings with learnable WordNet-relaxed text embeddings under a hierarchical loss. By avoiding direct use of exact class names and instead using semantically richer WordNet expansions, we mitigate label leakage and introduce more diverse textual cues. Experiments show that this strategy significantly boosts student performance, whereas noisy or overly precise text embeddings hinder distillation efficiency. Interpretability analyses confirm that WordNet-relaxed prompts encourage heavier reliance on visual features over textual shortcuts, while still effectively incorporating the newly introduced textual cues. Our method achieves state-of-the-art or second-best results on six public datasets, demonstrating its effectiveness in advancing crossmodal KD.
Continual self-supervised learning (CSSL) methods have gained increasing attention in remote sensing (RS) due to their capability to learn new tasks sequentially from continuous streams of unlabeled data. Existing CSSL methods, while learning new tasks, focus on preventing catastrophic forgetting. To this end, most of them use regularization strategies to retain knowledge of previous tasks. This reduces the model's ability to adapt to the data of new tasks (i.e., learning plasticity), which can degrade performance. To address this problem, in this paper, we propose a novel CSSL method that aims to learn tasks sequentially, while achieving high learning plasticity. To this end, the proposed method uses a knowledge distillation strategy with an integrated decoupling mechanism. The decoupling is achieved by first dividing the feature dimensions into task-common and task-specific parts. Then, the task-common features are forced to be correlated to ensure memory stability while the task-specific features are forced to be de-correlated facilitating the learning of new features. Experimental results show the effectiveness of the proposed method compared to CaSSLe, which is a widely used CSSL framework, with improvements of up to 1.12% in average accuracy and 2.33% in intransigence in a task-incremental scenario, and 1.24% in average accuracy and 2.01% in intransigence in a class-incremental scenario.
Multimodal Emotion Recognition in Conversations (MERC) identifies emotional states across text, audio and video, which is essential for intelligent dialogue systems and opinion analysis. Existing methods emphasize heterogeneous modal fusion directly for cross-modal integration, but often suffer from disorientation in multimodal learning due to modal heterogeneity and lack of instructive guidance. In this work, we propose SUMMER, a novel heterogeneous multimodal integration framework leveraging Mixture of Experts with Hierarchical Cross-modal Fusion and Interactive Knowledge Distillation. Key components include a Sparse Dynamic Mixture of Experts (SDMoE) for capturing dynamic token-wise interactions, a Hierarchical Cross-Modal Fusion (HCMF) for effective fusion of heterogeneous modalities, and Interactive Knowledge Distillation (IKD), which uses a pre-trained unimodal teacher to guide multimodal fusion in latent and logit spaces. Experiments on IEMOCAP and MELD show SUMMER outperforms state-of-the-art methods, particularly in recognizing minority and semantically similar emotions.
Pedestrian trajectory forecasting is crucial in various applications such as autonomous driving and mobile robot navigation. In such applications, camera-based perception enables the extraction of additional modalities (human pose, text) to enhance prediction accuracy. Indeed, we find that textual descriptions play a crucial role in integrating additional modalities into a unified understanding. However, online extraction of text requires the use of VLM, which may not be feasible for resource-constrained systems. To address this challenge, we propose a multi-modal knowledge distillation framework: a student model with limited modality is distilled from a teacher model trained with full range of modalities. The comprehensive knowledge of a teacher model trained with trajectory, human pose, and text is distilled into a student model using only trajectory or human pose as a sole supplement. In doing so, we separately distill the core locomotion insights from intra-agent multi-modality and inter-agent interaction. Our generalizable framework is validated with two state-of-the-art models across three datasets on both ego-view (JRDB, SIT) and BEV-view (ETH/UCY) setups, utilizing both annotated and VLM-generated text captions. Distilled student models show consistent improvement in all prediction metrics for both full and instantaneous observations, improving up to ~13%. The code is available at https://github.com/Jaewoo97/KDTF.
Growing data privacy demands, driven by regulations like GDPR and CCPA, require machine unlearning methods capable of swiftly removing the influence of specific training points. Although verified approaches like SISA, using data slicing and checkpointing, achieve efficient unlearning for single models by reverting to intermediate states, these methods struggle in teacher-student knowledge distillation settings. Unlearning in the teacher typically forces costly, complete student retraining due to pervasive information propagation during distillation. Our primary contribution is PURGE (Partitioned Unlearning with Retraining Guarantee for Ensembles), a novel framework integrating verified unlearning with distillation. We introduce constituent mapping and an incremental multi-teacher strategy that partitions the distillation process, confines each teacher constituent's impact to distinct student data subsets, and crucially maintains data isolation. The PURGE framework substantially reduces retraining overhead, requiring only partial student updates when teacher-side unlearning occurs. We provide both theoretical analysis, quantifying significant speed-ups in the unlearning process, and empirical validation on multiple datasets, demonstrating that PURGE achieves these efficiency gains while maintaining student accuracy comparable to standard baselines.
Self-supervised monocular depth estimation (SSMDE) has gained attention in the field of deep learning as it estimates depth without requiring ground truth depth maps. This approach typically uses a photometric consistency loss between a synthesized image, generated from the estimated depth, and the original image, thereby reducing the need for extensive dataset acquisition. However, the conventional photometric consistency loss relies on the Lambertian assumption, which often leads to significant errors when dealing with reflective surfaces that deviate from this model. To address this limitation, we propose a novel framework that incorporates intrinsic image decomposition into SSMDE. Our method synergistically trains for both monocular depth estimation and intrinsic image decomposition. The accurate depth estimation facilitates multi-image consistency for intrinsic image decomposition by aligning different view coordinate systems, while the decomposition process identifies reflective areas and excludes corrupted gradients from the depth training process. Furthermore, our framework introduces a pseudo-depth generation and knowledge distillation technique to further enhance the performance of the student model across both reflective and non-reflective surfaces. Comprehensive evaluations on multiple datasets show that our approach significantly outperforms existing SSMDE baselines in depth prediction, especially on reflective surfaces.
Recent years have witnessed the rapid progression of deep learning, pushing us closer to the realization of AGI (Artificial General Intelligence). Probabilistic modeling is critical to many of these advancements, which provides a foundational framework for capturing data distributions. However, as the scale and complexity of AI applications grow, traditional probabilistic modeling faces escalating challenges, such as high-dimensional parameter spaces, heterogeneous data sources, and evolving real-world requirements often render classical approaches insufficiently flexible. This paper proposes a novel concept, Probability Engineering, which treats the already-learned probability distributions within deep learning as engineering artifacts. Rather than merely fitting or inferring distributions, we actively modify and reinforce them to better address the diverse and evolving demands of modern AI. Specifically, Probability Engineering introduces novel techniques and constraints to refine existing probability distributions, improving their robustness, efficiency, adaptability, or trustworthiness. We showcase this paradigm through a series of applications spanning Bayesian deep learning, Edge AI (including federated learning and knowledge distillation), and Generative AI (such as text-to-image generation with diffusion models and high-quality text generation with large language models). These case studies demonstrate how probability distributions once treated as static objects can be engineered to meet the diverse and evolving requirements of large-scale, data-intensive, and trustworthy AI systems. By systematically expanding and strengthening the role of probabilistic modeling, Probability Engineering paves the way for more robust, adaptive, efficient, and trustworthy deep learning solutions in today's fast-growing AI era.
The modernization of smart farming is a way to improve agricultural production efficiency, and improve the agricultural production environment. Although many large models have achieved high accuracy in the task of object recognition and segmentation, they cannot really be put into use in the farming industry due to their own poor interpretability and limitations in computational volume. In this paper, we built AnYue Shelduck Dateset, which contains a total of 1951 Shelduck datasets, and performed target detection and segmentation annotation with the help of professional annotators. Based on AnYue ShelduckDateset, this paper describes DuckProcessing, an efficient and powerful module for duck identification based on real shelduckfarms. First of all, using the YOLOv8 module designed to divide the mahjong between them, Precision reached 98.10%, Recall reached 96.53% and F1 score reached 0.95 on the test set. Again using the DuckSegmentation segmentation model, DuckSegmentation reached 96.43% mIoU. Finally, the excellent DuckSegmentation was used as the teacher model, and through knowledge distillation, Deeplabv3 r50 was used as the student model, and the final student model achieved 94.49% mIoU on the test set. The method provides a new way of thinking in practical sisal duck smart farming.