Vision-Language Models (VLMs) bring powerful understanding and reasoning capabilities to multimodal tasks. Meanwhile, the great need for capable aritificial intelligence on mobile devices also arises, such as the AI assistant software. Some efforts try to migrate VLMs to edge devices to expand their application scope. Simplifying the model structure is a common method, but as the model shrinks, the trade-off between performance and size becomes more and more difficult. Knowledge distillation (KD) can help models improve comprehensive capabilities without increasing size or data volume. However, most of the existing large model distillation techniques only consider applications on single-modal LLMs, or only use teachers to create new data environments for students. None of these methods take into account the distillation of the most important cross-modal alignment knowledge in VLMs. We propose a method called Align-KD to guide the student model to learn the cross-modal matching that occurs at the shallow layer. The teacher also helps student learn the projection of vision token into text embedding space based on the focus of text. Under the guidance of Align-KD, the 1.7B MobileVLM V2 model can learn rich knowledge from the 7B teacher model with light design of training loss, and achieve an average score improvement of 2.0 across 6 benchmarks under two training subsets respectively. Code is available at: https://github.com/fqhank/Align-KD.
The use of artificial intelligence (AI) in automated disease classification significantly reduces healthcare costs and improves the accessibility of services. However, this transformation has given rise to concerns about the fairness of AI, which disproportionately affects certain groups, particularly patients from underprivileged populations. Recently, a number of methods and large-scale datasets have been proposed to address group performance disparities. Although these methods have shown effectiveness in disease classification tasks, they may fall short in ensuring fair prediction of disease progression, mainly because of limited longitudinal data with diverse demographics available for training a robust and equitable prediction model. In this paper, we introduce TransFair to enhance demographic fairness in progression prediction for ocular diseases. TransFair aims to transfer a fairness-enhanced disease classification model to the task of progression prediction with fairness preserved. Specifically, we train a fair EfficientNet, termed FairEN, equipped with a fairness-aware attention mechanism using extensive data for ocular disease classification. Subsequently, this fair classification model is adapted to a fair progression prediction model through knowledge distillation, which aims to minimize the latent feature distances between the classification and progression prediction models. We evaluate FairEN and TransFair for fairness-enhanced ocular disease classification and progression prediction using both two-dimensional (2D) and 3D retinal images. Extensive experiments and comparisons with models with and without considering fairness learning show that TransFair effectively enhances demographic equity in predicting ocular disease progression.
Large language models (LLMs) have demonstrated remarkable capabilities, but their adoption is limited by high computational costs during inference. While increasing parameter counts enhances accuracy, it also widens the gap between state-of-the-art capabilities and practical deployability. We present Puzzle, a framework to accelerate LLM inference on specific hardware while preserving their capabilities. Through an innovative application of neural architecture search (NAS) at an unprecedented scale, Puzzle systematically optimizes models with tens of billions of parameters under hardware constraints. Our approach utilizes blockwise local knowledge distillation (BLD) for parallel architecture exploration and employs mixed-integer programming for precise constraint optimization. We demonstrate the real-world impact of our framework through Llama-3.1-Nemotron-51B-Instruct (Nemotron-51B), a publicly available model derived from Llama-3.1-70B-Instruct. Nemotron-51B achieves a 2.17x inference throughput speedup, fitting on a single NVIDIA H100 GPU while preserving 98.4% of the original model's capabilities. Nemotron-51B currently stands as the most accurate language model capable of inference on a single GPU with large batch sizes. Remarkably, this transformation required just 45B training tokens, compared to over 15T tokens used for the 70B model it was derived from. This establishes a new paradigm where powerful models can be optimized for efficient deployment with only negligible compromise of their capabilities, demonstrating that inference performance, not parameter count alone, should guide model selection. With the release of Nemotron-51B and the presentation of the Puzzle framework, we provide practitioners immediate access to state-of-the-art language modeling capabilities at significantly reduced computational costs.
Product bundling aims to organize a set of thematically related items into a combined bundle for shipment facilitation and item promotion. To increase the exposure of fresh or overstocked products, sellers typically bundle these items with popular products for inventory clearance. This specific task can be formulated as a long-tail product bundling scenario, which leverages the user-item interactions to define the popularity of each item. The inherent popularity bias in the pre-extracted user feedback features and the insufficient utilization of other popularity-independent knowledge may force the conventional bundling methods to find more popular items, thereby struggling with this long-tail bundling scenario. Through intuitive and empirical analysis, we navigate the core solution for this challenge, which is maximally mining the popularity-free features and effectively incorporating them into the bundling process. To achieve this, we propose a Distilled Modality-Oriented Knowledge Transfer framework (DieT) to effectively counter the popularity bias misintroduced by the user feedback features and adhere to the original intent behind the real-world bundling behaviors. Specifically, DieT first proposes the Popularity-free Collaborative Distribution Modeling module (PCD) to capture the popularity-independent information from the bundle-item view, which is proven most effective in the long-tail bundling scenario to enable the directional information transfer. With the tailored Unbiased Bundle-aware Knowledge Transferring module (UBT), DieT can highlight the significance of popularity-free features while mitigating the negative effects of user feedback features in the long-tail scenario via the knowledge distillation paradigm. Extensive experiments on two real-world datasets demonstrate the superiority of DieT over a list of SOTA methods in the long-tail bundling scenario.
Zero-shot slot filling is a well-established subtask of Natural Language Understanding (NLU). However, most existing methods primarily focus on single-turn text data, overlooking the unique complexities of conversational dialogue. Conversational data is highly dynamic, often involving abrupt topic shifts, interruptions, and implicit references that make it difficult to directly apply zero-shot slot filling techniques, even with the remarkable capabilities of large language models (LLMs). This paper addresses these challenges by proposing strategies for automatic data annotation with slot induction and black-box knowledge distillation (KD) from a teacher LLM to a smaller model, outperforming vanilla LLMs on internal datasets by 26% absolute increase in F1 score. Additionally, we introduce an efficient system architecture for call center product settings that surpasses off-the-shelf extractive models by 34% relative F1 score, enabling near real-time inference on dialogue streams with higher accuracy, while preserving low latency.
Knowledge distillation (KD) is the de facto standard for compressing large-scale models into smaller ones. Prior works have explored ever more complex KD strategies involving different objective functions, teacher-ensembles, and weight inheritance. In this work we explore an alternative, yet simple approach -- active data curation as effective distillation for contrastive multimodal pretraining. Our simple online batch selection method, ACID, outperforms strong KD baselines across various model-, data- and compute-configurations. Further, we find such an active data curation strategy to in fact be complementary to standard KD, and can be effectively combined to train highly performant inference-efficient models. Our simple and scalable pretraining framework, ACED, achieves state-of-the-art results across 27 zero-shot classification and retrieval tasks with upto 11% less inference FLOPs. We further demonstrate that our ACED models yield strong vision-encoders for training generative multimodal models in the LiT-Decoder setting, outperforming larger vision encoders for image-captioning and visual question-answering tasks.
Effectively utilizing extensive unlabeled high-density EEG data to improve performance in scenarios with limited labeled low-density EEG data presents a significant challenge. In this paper, we address this by framing it as a graph transfer learning and knowledge distillation problem. We propose a Unified Pre-trained Graph Contrastive Masked Autoencoder Distiller, named EEG-DisGCMAE, to bridge the gap between unlabeled/labeled and high/low-density EEG data. To fully leverage the abundant unlabeled EEG data, we introduce a novel unified graph self-supervised pre-training paradigm, which seamlessly integrates Graph Contrastive Pre-training and Graph Masked Autoencoder Pre-training. This approach synergistically combines contrastive and generative pre-training techniques by reconstructing contrastive samples and contrasting the reconstructions. For knowledge distillation from high-density to low-density EEG data, we propose a Graph Topology Distillation loss function, allowing a lightweight student model trained on low-density data to learn from a teacher model trained on high-density data, effectively handling missing electrodes through contrastive distillation. To integrate transfer learning and distillation, we jointly pre-train the teacher and student models by contrasting their queries and keys during pre-training, enabling robust distillers for downstream tasks. We demonstrate the effectiveness of our method on four classification tasks across two clinical EEG datasets with abundant unlabeled data and limited labeled data. The experimental results show that our approach significantly outperforms contemporary methods in both efficiency and accuracy.
Reverse thinking plays a crucial role in human reasoning. Humans can reason not only from a problem to a solution but also in reverse, i.e., start from the solution and reason towards the problem. This often enhances overall reasoning performance as it enables consistency checks between their forward and backward thinking. To enable Large Language Models (LLMs) to perform reverse thinking, we introduce Reverse-Enhanced Thinking (RevThink), a framework composed of data augmentation and learning objectives. In RevThink, we augment the dataset by collecting structured forward-backward reasoning from a teacher model, consisting of: (1) the original question, (2) forward reasoning, (3) backward question, and (4) backward reasoning. We then employ three objectives to train a smaller student model in a multi-task learning fashion: (a) generate forward reasoning from a question, (b) generate a backward question from a question, and (c) generate backward reasoning from the backward question. Experiments across 12 datasets covering commonsense, math, and logical reasoning show an average 13.53% improvement over the student model's zero-shot performance and a 6.84% improvement over the strongest knowledge distillation baselines. Moreover, our method demonstrates sample efficiency -- using only 10% of the correct forward reasoning from the training data, it outperforms a standard fine-tuning method trained on 10x more forward reasoning. RevThink also exhibits strong generalization to out-of-distribution held-out datasets.
Low-resolution fine-grained image classification has recently made significant progress, largely thanks to the super-resolution techniques and knowledge distillation methods. However, these approaches lead to an exponential increase in the number of parameters and computational complexity of models. In order to solve this problem, in this letter, we propose a Vision Mamba Distillation (ViMD) approach to enhance the effectiveness and efficiency of low-resolution fine-grained image classification. Concretely, a lightweight super-resolution vision Mamba classification network (SRVM-Net) is proposed to improve its capability for extracting visual features by redesigning the classification sub-network with Mamba modeling. Moreover, we design a novel multi-level Mamba knowledge distillation loss boosting the performance, which can transfer prior knowledge obtained from a High-resolution Vision Mamba classification Network (HRVM-Net) as a teacher into the proposed SRVM-Net as a student. Extensive experiments on seven public fine-grained classification datasets related to benchmarks confirm our ViMD achieves a new state-of-the-art performance. While having higher accuracy, ViMD outperforms similar methods with fewer parameters and FLOPs, which is more suitable for embedded device applications. Code is available at https://github.com/boa2004plaust/ViMD.
The forest serves as the most significant terrestrial carbon stock mechanism, effectively reducing atmospheric CO$_2$ concentrations and mitigating climate change. Remote sensing provides high data accuracy and enables large-scale observations. Optical images facilitate long-term monitoring, which is crucial for future carbon stock estimation studies. This study focuses on Huize County, Qujing City, Yunnan Province, China, utilizing GF-1 WFV satellite imagery. The KD-VGG and KD-UNet modules were introduced for initial feature extraction, and the improved implicit diffusion model (IIDM) was proposed. The results showed: (1) The VGG module improved initial feature extraction, improving accuracy, and reducing inference time with optimized model parameters. (2) The Cross-attention + MLPs module enabled effective feature fusion, establishing critical relationships between global and local features, achieving high-accuracy estimation. (3) The IIDM model, a novel contribution, demonstrated the highest estimation accuracy with an RMSE of 12.17\%, significantly improving by 41.69\% to 42.33\% compared to the regression model. In carbon stock estimation, the generative model excelled in extracting deeper features, significantly outperforming other models, demonstrating the feasibility of AI-generated content in quantitative remote sensing. The 16-meter resolution estimates provide a robust basis for tailoring forest carbon sink regulations, enhancing regional carbon stock management.
Knowledge distillation (KD) has become a widely adopted approach for compressing large language models (LLMs) to reduce computational costs and memory footprints. However, the availability of complex teacher models is a prerequisite for running most KD pipelines. Thus, the traditional KD procedure can be unachievable or budget-unfriendly, particularly when relying on commercial LLMs like GPT4. In this regard, Self-distillation (SelfD) emerges as an advisable alternative, enabling student models to learn without teachers' guidance. Nonetheless, existing SelfD approaches for LMs often involve architectural modifications, assuming the models are open-source, which may not always be practical. In this work, we introduce a model-agnostic and task-agnostic method named dynamic SelfD from the previous minibatch (DynSDPB), which realizes current iterations' distillation from the last ones' generated logits. Additionally, to address prediction inaccuracies during the early iterations, we dynamically adjust the distillation influence and temperature values to enhance the adaptability of fine-tuning. Furthermore, DynSDPB is a novel fine-tuning policy that facilitates the seamless integration of existing self-correction and self-training techniques for small language models (SLMs) because they all require updating SLMs' parameters. We demonstrate the superior performance of DynSDPB on both encoder-only LMs (e.g., BERT model families) and decoder-only LMs (e.g., LLaMA model families), validating its effectiveness across natural language understanding (NLU) and natural language generation (NLG) benchmarks.
Data-Free Knowledge Distillation (DFKD) is an advanced technique that enables knowledge transfer from a teacher model to a student model without relying on original training data. While DFKD methods have achieved success on smaller datasets like CIFAR10 and CIFAR100, they encounter challenges on larger, high-resolution datasets such as ImageNet. A primary issue with previous approaches is their generation of synthetic images at high resolutions (e.g., $224 \times 224$) without leveraging information from real images, often resulting in noisy images that lack essential class-specific features in large datasets. Additionally, the computational cost of generating the extensive data needed for effective knowledge transfer can be prohibitive. In this paper, we introduce MUlti-reSolution data-freE (MUSE) to address these limitations. MUSE generates images at lower resolutions while using Class Activation Maps (CAMs) to ensure that the generated images retain critical, class-specific features. To further enhance model diversity, we propose multi-resolution generation and embedding diversity techniques that strengthen latent space representations, leading to significant performance improvements. Experimental results demonstrate that MUSE achieves state-of-the-art performance across both small- and large-scale datasets, with notable performance gains of up to two digits in nearly all ImageNet and subset experiments. Code is available at https://github.com/tmtuan1307/muse.
Vision-language foundation models, such as CLIP, have shown unprecedented zero-shot performance across a wide range of tasks. Nevertheless, these models may be unreliable under distributional shifts, as their performance is significantly degraded. In this work, we explore how to efficiently leverage class text information to mitigate these distribution drifts encountered by large pre-trained vision-language models (VLMs) during test-time inference. In particular, we propose to generate pseudo-labels for the test-time samples by exploiting generic class text embeddings as fixed centroids of a label assignment problem, which is efficiently solved with Optimal Transport. Furthermore, the proposed adaptation method (CLIP-OT) integrates a multiple template knowledge distillation approach, which replicates multi-view contrastive learning strategies in unsupervised representation learning but without incurring additional computational complexity. Extensive experiments on multiple popular test-time adaptation benchmarks presenting diverse complexity empirically show the superiority of CLIP-OT, achieving performance gains of up to 7% over recent state-of-the-art methods, yet being computationally and memory efficient.
One of the difficulties imposed on the manipulation of deformable objects is their characterization and the detection of representative keypoints for the purpose of manipulation. A keen interest was manifested by researchers in the last decade to characterize and manipulate deformable objects of non-fluid nature, such as clothes and ropes. Even though several propositions were made in the regard of object characterization, however researchers were always confronted with the need of pixel-level information of the object through images to extract relevant information. This usually is accomplished by means of segmentation networks trained on manually labeled data for this purpose. In this paper, we address the subject of characterizing weld pool to define stable features that serve as information for further motion control objectives. We achieve this by employing different pipelines. The first one consists of characterizing fluid deformable objects through the use of a generative model that is trained using a teacher-student framework. And in the second one we leverage foundation models by using them as teachers to characterize the object in the image, without the need of any pre-training and any dataset. The performance of knowledge distillation from foundation models into a smaller generative model shows prominent results in the characterization of deformable objects. The student network was capable of learning to retrieve the keypoitns of the object with an error of 13.4 pixels. And the teacher was evaluated based on its capacities to retrieve pixel level information represented by the object mask, with a mean Intersection Over Union (mIoU) of 75.26%.
Embedding models have become essential tools in both natural language processing and computer vision, enabling efficient semantic search, recommendation, clustering, and more. However, the high memory and computational demands of full-precision embeddings pose challenges for deployment in resource-constrained environments, such as real-time recommendation systems. In this work, we propose a novel finetuning framework to ternary-weight embedding models, which reduces memory and computational overhead while maintaining high performance. To apply ternarization to pre-trained embedding models, we introduce self-taught knowledge distillation to finalize the ternary-weights of the linear layers. With extensive experiments on public text and vision datasets, we demonstrated that without sacrificing effectiveness, the ternarized model consumes low memory usage and has low latency in the inference stage with great efficiency. In practical implementations, embedding models are typically integrated with Approximate Nearest Neighbor (ANN) search. Our experiments combining ternary embedding with ANN search yielded impressive improvement in both accuracy and computational efficiency. The repository is available at here.
Knowledge distillation approaches are model compression techniques, with the goal of training a highly performant student model by using a teacher network that is larger or contains a different inductive bias. These approaches are particularly useful when applied to large computer vision foundation models, which can be compressed into smaller variants that retain desirable properties such as improved robustness. This paper presents a label-free knowledge distillation approach called Teacher in the Middle (TinTeM), which improves on previous methods by learning an approximately orthogonal mapping from the latent space of the teacher to the student network. This produces a more faithful student, which better replicates the behavior of the teacher network across a range of benchmarks testing model robustness, generalisability and out-of-distribution detection. It is further shown that knowledge distillation with TinTeM on task specific datasets leads to more accurate models with greater generalisability and OOD detection performance, and that this technique provides a competitive pathway for training highly performant lightweight models on small datasets.
Click-through rate (CTR) prediction plays a critical role in recommender systems and web searches. While many existing methods utilize ensemble learning to improve model performance, they typically limit the ensemble to two or three sub-networks, with little exploration of larger ensembles. In this paper, we investigate larger ensemble networks and find three inherent limitations in commonly used ensemble learning method: (1) performance degradation with more networks; (2) sharp decline and high variance in sub-network performance; (3) large discrepancies between sub-network and ensemble predictions. To simultaneously address the above limitations, this paper investigates potential solutions from the perspectives of Knowledge Distillation (KD) and Deep Mutual Learning (DML). Based on the empirical performance of these methods, we combine them to propose a novel model-agnostic Ensemble Knowledge Transfer Framework (EKTF). Specifically, we employ the collective decision-making of the students as an abstract teacher to guide each student (sub-network) towards more effective learning. Additionally, we encourage mutual learning among students to enable knowledge acquisition from different views. To address the issue of balancing the loss hyperparameters, we design a novel examination mechanism to ensure tailored teaching from teacher-to-student and selective learning in peer-to-peer. Experimental results on five real-world datasets demonstrate the effectiveness and compatibility of EKTF. The code, running logs, and detailed hyperparameter configurations are available at: https://github.com/salmon1802/EKTF.
Sentence-level embedding is essential for various tasks that require understanding natural language. Many studies have explored such embeddings for high-resource languages like English. However, low-resource languages like Bengali (a language spoken by almost two hundred and thirty million people) are still under-explored. This work introduces two lightweight sentence transformers for the Bangla language, leveraging a novel cross-lingual knowledge distillation approach. This method distills knowledge from a pre-trained, high-performing English sentence transformer. Proposed models are evaluated across multiple downstream tasks, including paraphrase detection, semantic textual similarity (STS), and Bangla hate speech detection. The new method consistently outperformed existing Bangla sentence transformers. Moreover, the lightweight architecture and shorter inference time make the models highly suitable for deployment in resource-constrained environments, making them valuable for practical NLP applications in low-resource languages.
Recent advancements in vision-language models (VLMs), such as CLIP, have demonstrated substantial success in self-supervised representation learning for vision tasks. However, effectively adapting VLMs to downstream applications remains challenging, as their accuracy often depends on time-intensive and expertise-demanding prompt engineering, while full model fine-tuning is costly. This is particularly true for biomedical images, which, unlike natural images, typically suffer from limited annotated datasets, unintuitive image contrasts, and nuanced visual features. Recent prompt learning techniques, such as Context Optimization (CoOp) intend to tackle these issues, but still fall short in generalizability. Meanwhile, explorations in prompt learning for biomedical image analysis are still highly limited. In this work, we propose BiomedCoOp, a novel prompt learning framework that enables efficient adaptation of BiomedCLIP for accurate and highly generalizable few-shot biomedical image classification. Our approach achieves effective prompt context learning by leveraging semantic consistency with average prompt ensembles from Large Language Models (LLMs) and knowledge distillation with a statistics-based prompt selection strategy. We conducted comprehensive validation of our proposed framework on 11 medical datasets across 9 modalities and 10 organs against existing state-of-the-art methods, demonstrating significant improvements in both accuracy and generalizability. The code will be publicly available at https://github.com/HealthX-Lab/BiomedCoOp.
Pretrained models have revolutionized deep learning by enabling significant performance improvements across a wide range of tasks, leveraging large-scale, pre-learned knowledge representations. However, deploying these models in real-world multi-task learning (MTL) scenarios poses substantial challenges, primarily due to high computational costs and inefficiencies in inference. Traditional approaches such as pruning, quantization, and knowledge distillation have been explored to mitigate these issues, but they often fall short in fully addressing the complexities of multi-task environments. This paper introduces \textbf{\underline{S}}elective \textbf{\underline{T}}ask \textbf{\underline{A}}rithmetic \underline{\textbf{(STA)}}, a training-free framework designed to enhance multi-task performance through task-specific parameter fusion. STA addresses three key challenges: (i) \textbf{Parameter importance diversity: } Recognizing that different tasks relie on distinct parameters, STA employs a loss-sensitive parameter importance metric derived from a first-order Taylor expansion to accurately measure the importance of parameters for each task. (ii) \textbf{Over-reliance on hyperparameter tuning: }By enhancing the sparsity of task vectors through parameter importance metrics, STA reduces the need for extensive hyperparameter tuning, thereby improving the generalization and robustness of the model. (iii) \textbf{Neglect of other abilities in task arithmetic: } Previous works have largely overlooked the potential for more precise task forgetting. STA leverages its parameter importance metric to achieve more controlled and effective task forgetting, minimizing the impact of noisy elements that can degrade model performance. Experimental results demonstrate that STA achieves superior multi-task performance across benchmarks and excellent performance in task forgetting.