Substantial efforts have been devoted to alleviating the impact of the long-tailed class distribution in federated learning. In this work, we observe an interesting phenomenon that weak classes consistently exist even for class-balanced learning. These weak classes, different from the minority classes in the previous works, are inherent to data and remain fairly consistent for various network structures and learning paradigms. The inherent inter-class accuracy discrepancy can reach over 36.9% for federated learning on the FashionMNIST and CIFAR-10 datasets, even when the class distribution is balanced both globally and locally. In this study, we empirically analyze the potential reason for this phenomenon. Furthermore, a class-specific partial knowledge distillation method is proposed to improve the model's classification accuracy for weak classes. In this approach, knowledge transfer is initiated upon the occurrence of specific misclassifications within certain weak classes. Experimental results show that the accuracy of weak classes can be improved by 10.7%, reducing the inherent interclass discrepancy effectively.
Large pre-trained Vision-Language Models (VLMs) such as Contrastive Language-Image Pre-Training (CLIP) have been shown to be susceptible to adversarial attacks, raising concerns about their deployment in safety-critical scenarios like autonomous driving and medical diagnosis. One promising approach for improving the robustness of pre-trained VLMs is Adversarial Prompt Tuning (APT), which combines adversarial training with prompt tuning. However, existing APT methods are mostly single-modal methods that design prompt(s) for only the visual or textual modality, limiting their effectiveness in either robustness or clean accuracy. In this work, we propose a novel method called Adversarial Prompt Distillation (APD) that combines APT with knowledge distillation to boost the adversarial robustness of CLIP. Specifically, APD is a bimodal method that adds prompts for both the visual and textual modalities while leveraging a cleanly pre-trained teacher CLIP model to distill and boost the performance of the student CLIP model on downstream tasks. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our APD over the current state-of-the-art APT methods in terms of both natural and adversarial performances. The effectiveness of our APD method validates the possibility of using a non-robust teacher to improve the generalization and robustness of VLMs.
This paper presents a critical examination of current approaches to replicating OpenAI's O1 model capabilities, with particular focus on the widespread but often undisclosed use of knowledge distillation techniques. While our previous work explored the fundamental technical path to O1 replication, this study reveals how simple distillation from O1's API, combined with supervised fine-tuning, can achieve superior performance on complex mathematical reasoning tasks. Through extensive experiments, we show that a base model fine-tuned on simply tens of thousands of samples O1-distilled long-thought chains outperforms O1-preview on the American Invitational Mathematics Examination (AIME) with minimal technical complexity. Moreover, our investigation extends beyond mathematical reasoning to explore the generalization capabilities of O1-distilled models across diverse tasks: hallucination, safety and open-domain QA. Notably, despite training only on mathematical problem-solving data, our models demonstrated strong generalization to open-ended QA tasks and became significantly less susceptible to sycophancy after fine-tuning. We deliberately make this finding public to promote transparency in AI research and to challenge the current trend of obscured technical claims in the field. Our work includes: (1) A detailed technical exposition of the distillation process and its effectiveness, (2) A comprehensive benchmark framework for evaluating and categorizing O1 replication attempts based on their technical transparency and reproducibility, (3) A critical discussion of the limitations and potential risks of over-relying on distillation approaches, our analysis culminates in a crucial bitter lesson: while the pursuit of more capable AI systems is important, the development of researchers grounded in first-principles thinking is paramount.
We present our submission to the BabyLM challenge, aiming to push the boundaries of data-efficient language model pretraining. Our method builds upon deep mutual learning, introducing a student model search for diverse initialization. We address the limitation of treating students equally by formulating weighted mutual learning as a bi-level optimization problem. The inner loop learns compact students through online distillation, while the outer loop optimizes weights for better knowledge distillation from diverse students. This dynamic weighting strategy eliminates the need for a teacher model, reducing computational requirements. Our evaluations show that teacher-less methods can match or surpass teacher-supervised approaches.
Recent breakthroughs in large foundation models have enabled the possibility of transferring knowledge pre-trained on vast datasets to domains with limited data availability. Agriculture is one of the domains that lacks sufficient data. This study proposes a framework to train effective, domain-specific, small models from foundation models without manual annotation. Our approach begins with SDM (Segmentation-Description-Matching), a stage that leverages two foundation models: SAM2 (Segment Anything in Images and Videos) for segmentation and OpenCLIP (Open Contrastive Language-Image Pretraining) for zero-shot open-vocabulary classification. In the second stage, a novel knowledge distillation mechanism is utilized to distill compact, edge-deployable models from SDM, enhancing both inference speed and perception accuracy. The complete method, termed SDM-D (Segmentation-Description-Matching-Distilling), demonstrates strong performance across various fruit detection tasks object detection, semantic segmentation, and instance segmentation) without manual annotation. It nearly matches the performance of models trained with abundant labels. Notably, SDM-D outperforms open-set detection methods such as Grounding SAM and YOLO-World on all tested fruit detection datasets. Additionally, we introduce MegaFruits, a comprehensive fruit segmentation dataset encompassing over 25,000 images, and all code and datasets are made publicly available at https://github.com/AgRoboticsResearch/SDM-D.git.
Neural networks can learn spurious correlations in the data, often leading to performance disparity for underrepresented subgroups. Studies have demonstrated that the disparity is amplified when knowledge is distilled from a complex teacher model to a relatively "simple" student model. Prior work has shown that ensemble deep learning methods can improve the performance of the worst-case subgroups; however, it is unclear if this advantage carries over when distilling knowledge from an ensemble of teachers, especially when the teacher models are debiased. This study demonstrates that traditional ensemble knowledge distillation can significantly drop the performance of the worst-case subgroups in the distilled student model even when the teacher models are debiased. To overcome this, we propose Adaptive Group Robust Ensemble Knowledge Distillation (AGRE-KD), a simple ensembling strategy to ensure that the student model receives knowledge beneficial for unknown underrepresented subgroups. Leveraging an additional biased model, our method selectively chooses teachers whose knowledge would better improve the worst-performing subgroups by upweighting the teachers with gradient directions deviating from the biased model. Our experiments on several datasets demonstrate the superiority of the proposed ensemble distillation technique and show that it can even outperform classic model ensembles based on majority voting.
Invoices and receipts submitted by employees are visually rich documents (VRDs) with textual, visual and layout information. To protect against the risk of fraud and abuse, it is crucial for organizations to efficiently extract desired information from submitted receipts. This helps in the assessment of key factors such as appropriateness of the expense claim, adherence to spending and transaction policies, the validity of the receipt, as well as downstream anomaly detection at various levels. These documents are heterogeneous, with multiple formats and languages, uploaded with different image qualities, and often do not contain ground truth labels for the efficient training of models. In this paper we propose Task Aware Instruction-based Labelling (TAIL), a method for synthetic label generation in VRD corpuses without labels, and fine-tune a multimodal Visually Rich Document Understanding Model (VRDU) on TAIL labels using response-based knowledge distillation without using the teacher model's weights or training dataset to conditionally generate annotations in the appropriate format. Using a benchmark external dataset where ground truth labels are available, we demonstrate conditions under which our approach performs at par with Claude 3 Sonnet through empirical studies. We then show that the resulting model performs at par or better on the internal expense documents of a large multinational organization than state-of-the-art LMM (large multimodal model) Claude 3 Sonnet while being 85% less costly and ~5X faster, and outperforms layout-aware baselines by more than 10% in Average Normalized Levenshtein Similarity (ANLS) scores due to its ability to reason and extract information from rare formats. Finally, we illustrate the usage of our approach in overpayment prevention.
Spatial transcriptomics (ST) provides essential spatial context by mapping gene expression within tissue, enabling detailed study of cellular heterogeneity and tissue organization. However, aligning ST data with histology images poses challenges due to inherent spatial distortions and modality-specific variations. Existing methods largely rely on direct alignment, which often fails to capture complex cross-modal relationships. To address these limitations, we propose a novel framework that aligns gene and image features using a ranking-based alignment loss, preserving relative similarity across modalities and enabling robust multi-scale alignment. To further enhance the alignment's stability, we employ self-supervised knowledge distillation with a teacher-student network architecture, effectively mitigating disruptions from high dimensionality, sparsity, and noise in gene expression data. Extensive experiments on seven public datasets that encompass gene expression prediction, slide-level classification, and survival analysis demonstrate the efficacy of our method, showing improved alignment and predictive performance over existing methods.
Contrastive Language-Image Pre-training (CLIP) has attracted a surge of attention for its superior zero-shot performance and excellent transferability to downstream tasks. However, training such large-scale models usually requires substantial computation and storage, which poses barriers for general users with consumer-level computers. Motivated by this observation, in this paper we investigate how to achieve competitive performance on only one Nvidia RTX3090 GPU and with one terabyte for storing dataset. On one hand, we simplify the transformer block structure and combine Weight Inheritance with multi-stage Knowledge Distillation (WIKD), thereby reducing the parameters and improving the inference speed during training along with deployment. On the other hand, confronted with the convergence challenge posed by small dataset, we generate synthetic captions for each sample as data augmentation, and devise a novel Pair Matching (PM) loss to fully exploit the distinguishment among positive and negative image-text pairs. Extensive experiments demonstrate that our model can achieve a new state-of-the-art datascale-parameter-accuracy tradeoff, which could further popularize the CLIP model in the related research community.
Large Language Models (LLMs) demonstrate exceptional reasoning capabilities, often achieving state-of-the-art performance in various tasks. However, their substantial computational and memory demands, due to billions of parameters, hinder deployment in resource-constrained environments. A promising solution is knowledge distillation, where LLMs transfer reasoning capabilities to Small Language Models (SLMs, $\le$ 1B parameters), enabling wider deployment on low-resource devices. Existing methods primarily focus on generating high-quality reasoning rationales for distillation datasets but often neglect the critical role of data quantity and quality. To address these challenges, we propose a Feedback-Driven Distillation (FDD) framework to enhance SLMs' mathematical reasoning capabilities. In the initialization stage, a distillation dataset is constructed by prompting LLMs to pair mathematical problems with corresponding reasoning rationales. We classify problems into easy and hard categories based on SLM performance. For easy problems, LLMs generate more complex variations, while for hard problems, new questions of similar complexity are synthesized. In addition, we propose a multi-round distillation paradigm to iteratively enrich the distillation datasets, thereby progressively improving the mathematical reasoning abilities of SLMs. Experimental results demonstrate that our method can make SLMs achieve SOTA mathematical reasoning performance.
Heterogeneous Graph Neural Networks (HGNNs) have achieved promising results in various heterogeneous graph learning tasks, owing to their superiority in capturing the intricate relationships and diverse relational semantics inherent in heterogeneous graph structures. However, the neighborhood-fetching latency incurred by structure dependency in HGNNs makes it challenging to deploy for latency-constrained applications that require fast inference. Inspired by recent GNN-to-MLP knowledge distillation frameworks, we introduce HG2M and HG2M+ to combine both HGNN's superior performance and MLP's efficient inference. HG2M directly trains student MLPs with node features as input and soft labels from teacher HGNNs as targets, and HG2M+ further distills reliable and heterogeneous semantic knowledge into student MLPs through reliable node distillation and reliable meta-path distillation. Experiments conducted on six heterogeneous graph datasets show that despite lacking structural dependencies, HG2Ms can still achieve competitive or even better performance than HGNNs and significantly outperform vanilla MLPs. Moreover, HG2Ms demonstrate a 379.24$\times$ speedup in inference over HGNNs on the large-scale IGB-3M-19 dataset, showcasing their ability for latency-sensitive deployments.
This paper introduces WARLearn, a novel framework designed for adaptive representation learning in challenging and adversarial weather conditions. Leveraging the in-variance principal used in Barlow Twins, we demonstrate the capability to port the existing models initially trained on clear weather data to effectively handle adverse weather conditions. With minimal additional training, our method exhibits remarkable performance gains in scenarios characterized by fog and low-light conditions. This adaptive framework extends its applicability beyond adverse weather settings, offering a versatile solution for domains exhibiting variations in data distributions. Furthermore, WARLearn is invaluable in scenarios where data distributions undergo significant shifts over time, enabling models to remain updated and accurate. Our experimental findings reveal a remarkable performance, with a mean average precision (mAP) of 52.6% on unseen real-world foggy dataset (RTTS). Similarly, in low light conditions, our framework achieves a mAP of 55.7% on unseen real-world low light dataset (ExDark). Notably, WARLearn surpasses the performance of state-of-the-art frameworks including FeatEnHancer, Image Adaptive YOLO, DENet, C2PNet, PairLIE and ZeroDCE, by a substantial margin in adverse weather, improving the baseline performance in both foggy and low light conditions. The WARLearn code is available at https://github.com/ShubhamAgarwal12/WARLearn
An important aspect of deploying face recognition (FR) algorithms in real-world applications is their ability to learn new face identities from a continuous data stream. However, the online training of existing deep neural network-based FR algorithms, which are pre-trained offline on large-scale stationary datasets, encounter two major challenges: (I) catastrophic forgetting of previously learned identities, and (II) the need to store past data for complete retraining from scratch, leading to significant storage constraints and privacy concerns. In this paper, we introduce CLFace, a continual learning framework designed to preserve and incrementally extend the learned knowledge. CLFace eliminates the classification layer, resulting in a resource-efficient FR model that remains fixed throughout lifelong learning and provides label-free supervision to a student model, making it suitable for open-set face recognition during incremental steps. We introduce an objective function that employs feature-level distillation to reduce drift between feature maps of the student and teacher models across multiple stages. Additionally, it incorporates a geometry-preserving distillation scheme to maintain the orientation of the teacher model's feature embedding. Furthermore, a contrastive knowledge distillation is incorporated to continually enhance the discriminative power of the feature representation by matching similarities between new identities. Experiments on several benchmark FR datasets demonstrate that CLFace outperforms baseline approaches and state-of-the-art methods on unseen identities using both in-domain and out-of-domain datasets.
Knowledge distillation (KD) has been a popular and effective method for model compression. One important assumption of KD is that the teacher's original dataset will also be available when training the student. However, in situations such as continual learning and distilling large models trained on company-withheld datasets, having access to the original data may not always be possible. This leads practitioners towards utilizing other sources of supplemental data, which could yield mixed results. One must then ask: "what makes a good dataset for transferring knowledge from teacher to student?" Many would assume that only real in-domain imagery is viable, but is that the only option? In this work, we explore multiple possible surrogate distillation datasets and demonstrate that many different datasets, even unnatural synthetic imagery, can serve as a suitable alternative in KD. From examining these alternative datasets, we identify and present various criteria describing what makes a good dataset for distillation. Source code will be available in the future.
Super-resolution (SR) is a key technique for improving the visual quality of video content by increasing its spatial resolution while reconstructing fine details. SR has been employed in many applications including video streaming, where compressed low-resolution content is typically transmitted to end users and then reconstructed with a higher resolution and enhanced quality. To support real-time playback, it is important to implement fast SR models while preserving reconstruction quality; however most existing solutions, in particular those based on complex deep neural networks, fail to do so. To address this issue, this paper proposes a low-complexity SR method, RTSR, designed to enhance the visual quality of compressed video content, focusing on resolution up-scaling from a) 360p to 1080p and from b) 540p to 4K. The proposed approach utilizes a CNN-based network architecture, which was optimized for AV1 (SVT)-encoded content at various quantization levels based on a dual-teacher knowledge distillation method. This method was submitted to the AIM 2024 Video Super-Resolution Challenge, specifically targeting the Efficient/Mobile Real-Time Video Super-Resolution competition. It achieved the best trade-off between complexity and coding performance (measured in PSNR, SSIM and VMAF) among all six submissions. The code will be available soon.
Effective query-item relevance modeling is pivotal for enhancing user experience and safeguarding user satisfaction in e-commerce search systems. Recently, benefiting from the vast inherent knowledge, Large Language Model (LLM) approach demonstrates strong performance and long-tail generalization ability compared with previous neural-based specialized relevance learning methods. Though promising, current LLM-based methods encounter the following inadequacies in practice: First, the massive parameters and computational demands make it difficult to be deployed online. Second, distilling LLM models to online models is a feasible direction, but the LLM relevance modeling is a black box, and its rich intrinsic knowledge is difficult to extract and apply online. To improve the interpretability of LLM and boost the performance of online relevance models via LLM, we propose an Explainable LLM-driven Multi-dimensional Distillation framework for e-commerce relevance learning, which comprises two core components: (1) An Explainable LLM for relevance modeling (ELLM-rele), which decomposes the relevance learning into intermediate steps and models relevance learning as a Chain-of-Thought (CoT) reasoning, thereby enhancing both interpretability and performance of LLM. (2) A Multi-dimensional Knowledge Distillation (MKD) architecture that transfers the knowledge of ELLM-rele to current deployable interaction-based and representation-based student models from both the relevance score distribution and CoT reasoning aspects. Through distilling the probabilistic and CoT reasoning knowledge, MKD improves both the semantic interaction and long-tail generalization abilities of student models. Extensive offline evaluations and online experiments on Taobao search ad scene demonstrate that our proposed framework significantly enhances e-commerce relevance learning performance and user experience.
Learning a reward model (RM) from human preferences has been an important component in aligning large language models (LLMs). The canonical setup of learning RMs from pairwise preference data is rooted in the classic Bradley-Terry (BT) model that accepts binary feedback, i.e., the label being either Response 1 is better than Response 2, or the opposite. Such a setup inevitably discards potentially useful samples (such as "tied" between the two responses) and loses more fine-grained information (such as "slightly better"). In this paper, we propose a framework for learning RMs under ordinal feedback which generalizes the case of binary preference feedback to any arbitrary granularity. Specifically, we first identify a marginal unbiasedness condition, which generalizes the assumption of the BT model in the existing binary feedback setting. The condition validates itself via the sociological concept of the wisdom of the crowd. Under the condition, we develop a natural probability model for pairwise preference data under ordinal feedback and analyze its properties. We prove the statistical benefits of ordinal feedback in terms of reducing the Rademacher complexity compared to the case of binary feedback. The proposed learning objective and the theory also extend to hinge loss and direct policy optimization (DPO). In particular, the theoretical analysis may be of independent interest when applying to a seemingly unrelated problem of knowledge distillation to interpret the bias-variance trade-off therein. The framework also sheds light on writing guidance for human annotators. Our numerical experiments validate that fine-grained feedback leads to better reward learning for both in-distribution and out-of-distribution settings. Further experiments show that incorporating a certain proportion of samples with tied preference boosts RM learning.
In this work, we attempted to extend the thought and showcase a way forward for the Self-supervised Learning (SSL) learning paradigm by combining contrastive learning, self-distillation (knowledge distillation) and masked data modelling, the three major SSL frameworks, to learn a joint and coordinated representation. The proposed technique of SSL learns by the collaborative power of different learning objectives of SSL. Hence to jointly learn the different SSL objectives we proposed a new SSL architecture KDC-MAE, a complementary masking strategy to learn the modular correspondence, and a weighted way to combine them coordinately. Experimental results conclude that the contrastive masking correspondence along with the KD learning objective has lent a hand to performing better learning for multiple modalities over multiple tasks.
Toxicity identification in online multimodal environments remains a challenging task due to the complexity of contextual connections across modalities (e.g., textual and visual). In this paper, we propose a novel framework that integrates Knowledge Distillation (KD) from Large Visual Language Models (LVLMs) and knowledge infusion to enhance the performance of toxicity detection in hateful memes. Our approach extracts sub-knowledge graphs from ConceptNet, a large-scale commonsense Knowledge Graph (KG) to be infused within a compact VLM framework. The relational context between toxic phrases in captions and memes, as well as visual concepts in memes enhance the model's reasoning capabilities. Experimental results from our study on two hate speech benchmark datasets demonstrate superior performance over the state-of-the-art baselines across AU-ROC, F1, and Recall with improvements of 1.1%, 7%, and 35%, respectively. Given the contextual complexity of the toxicity detection task, our approach showcases the significance of learning from both explicit (i.e. KG) as well as implicit (i.e. LVLMs) contextual cues incorporated through a hybrid neurosymbolic approach. This is crucial for real-world applications where accurate and scalable recognition of toxic content is critical for creating safer online environments.
Federated Named Entity Recognition (FNER) boosts model training within each local client by aggregating the model updates of decentralized local clients, without sharing their private data. However, existing FNER methods assume fixed entity types and local clients in advance, leading to their ineffectiveness in practical applications. In a more realistic scenario, local clients receive new entity types continuously, while new local clients collecting novel data may irregularly join the global FNER training. This challenging setup, referred to here as Federated Incremental NER, renders the global model suffering from heterogeneous forgetting of old entity types from both intra-client and inter-client perspectives. To overcome these challenges, we propose a Local-Global Forgetting Defense (LGFD) model. Specifically, to address intra-client forgetting, we develop a structural knowledge distillation loss to retain the latent space's feature structure and a pseudo-label-guided inter-type contrastive loss to enhance discriminative capability over different entity types, effectively preserving previously learned knowledge within local clients. To tackle inter-client forgetting, we propose a task switching monitor that can automatically identify new entity types under privacy protection and store the latest old global model for knowledge distillation and pseudo-labeling. Experiments demonstrate significant improvement of our LGFD model over comparison methods.