In this paper, we present a new embedding model, called M3-Embedding, which is distinguished for its versatility in Multi-Linguality, Multi-Functionality, and Multi-Granularity. It can support more than 100 working languages, leading to new state-of-the-art performances on multi-lingual and cross-lingual retrieval tasks. It can simultaneously perform the three common retrieval functionalities of embedding model: dense retrieval, multi-vector retrieval, and sparse retrieval, which provides a unified model foundation for real-world IR applications. It is able to process inputs of different granularities, spanning from short sentences to long documents of up to 8192 tokens. The effective training of M3-Embedding involves the following technical contributions. We propose a novel self-knowledge distillation approach, where the relevance scores from different retrieval functionalities can be integrated as the teacher signal to enhance the training quality. We also optimize the batching strategy, enabling a large batch size and high training throughput to ensure the discriminativeness of embeddings. To the best of our knowledge, M3-Embedding is the first embedding model which realizes such a strong versatility. The model and code will be publicly available at https://github.com/FlagOpen/FlagEmbedding.
Knowledge Distillation (KD) has proven effective for compressing large teacher models into smaller student models. While it is well known that student models can achieve similar accuracies as the teachers, it has also been shown that they nonetheless often do not learn the same function. It is, however, often highly desirable that the student's and teacher's functions share similar properties such as basing the prediction on the same input features, as this ensures that students learn the 'right features' from the teachers. In this work, we explore whether this can be achieved by not only optimizing the classic KD loss but also the similarity of the explanations generated by the teacher and the student. Despite the idea being simple and intuitive, we find that our proposed 'explanation-enhanced' KD (e$^2$KD) (1) consistently provides large gains in terms of accuracy and student-teacher agreement, (2) ensures that the student learns from the teacher to be right for the right reasons and to give similar explanations, and (3) is robust with respect to the model architectures, the amount of training data, and even works with 'approximate', pre-computed explanations.
The recommendation of medication is a vital aspect of intelligent healthcare systems, as it involves prescribing the most suitable drugs based on a patient's specific health needs. Unfortunately, many sophisticated models currently in use tend to overlook the nuanced semantics of medical data, while only relying heavily on identities. Furthermore, these models face significant challenges in handling cases involving patients who are visiting the hospital for the first time, as they lack prior prescription histories to draw upon. To tackle these issues, we harness the powerful semantic comprehension and input-agnostic characteristics of Large Language Models (LLMs). Our research aims to transform existing medication recommendation methodologies using LLMs. In this paper, we introduce a novel approach called Large Language Model Distilling Medication Recommendation (LEADER). We begin by creating appropriate prompt templates that enable LLMs to suggest medications effectively. However, the straightforward integration of LLMs into recommender systems leads to an out-of-corpus issue specific to drugs. We handle it by adapting the LLMs with a novel output layer and a refined tuning loss function. Although LLM-based models exhibit remarkable capabilities, they are plagued by high computational costs during inference, which is impractical for the healthcare sector. To mitigate this, we have developed a feature-level knowledge distillation technique, which transfers the LLM's proficiency to a more compact model. Extensive experiments conducted on two real-world datasets, MIMIC-III and MIMIC-IV, demonstrate that our proposed model not only delivers effective results but also is efficient. To ease the reproducibility of our experiments, we release the implementation code online.
Humans are capable of acquiring new knowledge and transferring learned knowledge into different domains, incurring a small forgetting. The same ability, called Continual Learning, is challenging to achieve when operating with neural networks due to the forgetting affecting past learned tasks when learning new ones. This forgetting can be mitigated by replaying stored samples from past tasks, but a large memory size may be needed for long sequences of tasks; moreover, this could lead to overfitting on saved samples. In this paper, we propose a novel regularisation approach and a novel incremental classifier called, respectively, Margin Dampening and Cascaded Scaling Classifier. The first combines a soft constraint and a knowledge distillation approach to preserve past learned knowledge while allowing the model to learn new patterns effectively. The latter is a gated incremental classifier, helping the model modify past predictions without directly interfering with them. This is achieved by modifying the output of the model with auxiliary scaling functions. We empirically show that our approach performs well on multiple benchmarks against well-established baselines, and we also study each component of our proposal and how the combinations of such components affect the final results.
Intersection over Union (IoU) losses are surrogates that directly optimize the Jaccard index. Leveraging IoU losses as part of the loss function have demonstrated superior performance in semantic segmentation tasks compared to optimizing pixel-wise losses such as the cross-entropy loss alone. However, we identify a lack of flexibility in these losses to support vital training techniques like label smoothing, knowledge distillation, and semi-supervised learning, mainly due to their inability to process soft labels. To address this, we introduce Jaccard Metric Losses (JMLs), which are identical to the soft Jaccard loss in standard settings with hard labels but are fully compatible with soft labels. We apply JMLs to three prominent use cases of soft labels: label smoothing, knowledge distillation and semi-supervised learning, and demonstrate their potential to enhance model accuracy and calibration. Our experiments show consistent improvements over the cross-entropy loss across 4 semantic segmentation datasets (Cityscapes, PASCAL VOC, ADE20K, DeepGlobe Land) and 13 architectures, including classic CNNs and recent vision transformers. Remarkably, our straightforward approach significantly outperforms state-of-the-art knowledge distillation and semi-supervised learning methods. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.
Reducing hallucination of Large Language Models (LLMs) is imperative for use in the sciences, where reliability and reproducibility are crucial. However, LLMs inherently lack long-term memory, making it a nontrivial, ad hoc, and inevitably biased task to fine-tune them on domain-specific literature and data. Here we introduce LLaMP, a multimodal retrieval-augmented generation (RAG) framework of hierarchical reasoning-and-acting (ReAct) agents that can dynamically and recursively interact with computational and experimental data on Materials Project (MP) and run atomistic simulations via high-throughput workflow interface. Without fine-tuning, LLaMP demonstrates strong tool usage ability to comprehend and integrate various modalities of materials science concepts, fetch relevant data stores on the fly, process higher-order data (such as crystal structure and elastic tensor), and streamline complex tasks in computational materials and chemistry. We propose a simple metric combining uncertainty and confidence estimates to evaluate the self-consistency of responses by LLaMP and vanilla LLMs. Our benchmark shows that LLaMP effectively mitigates the intrinsic bias in LLMs, counteracting the errors on bulk moduli, electronic bandgaps, and formation energies that seem to derive from mixed data sources. We also demonstrate LLaMP's capability to edit crystal structures and run annealing molecular dynamics simulations using pre-trained machine-learning force fields. The framework offers an intuitive and nearly hallucination-free approach to exploring and scaling materials informatics, and establishes a pathway for knowledge distillation and fine-tuning other language models. Code and live demo are available at https://github.com/chiang-yuan/llamp
In learning-based functionality stealing, the attacker is trying to build a local model based on the victim's outputs. The attacker has to make choices regarding the local model's architecture, optimization method and, specifically for NLP models, subword vocabulary, such as BPE. On the machine translation task, we explore (1) whether the choice of the vocabulary plays a role in model stealing scenarios and (2) if it is possible to extract the victim's vocabulary. We find that the vocabulary itself does not have a large effect on the local model's performance. Given gray-box model access, it is possible to collect the victim's vocabulary by collecting the outputs (detokenized subwords on the output). The results of the minimum effect of vocabulary choice are important more broadly for black-box knowledge distillation.
We introduce TQCompressor, a novel method for neural network model compression with improved tensor decompositions. We explore the challenges posed by the computational and storage demands of pre-trained language models in NLP tasks and propose a permutation-based enhancement to Kronecker decomposition. This enhancement makes it possible to reduce loss in model expressivity which is usually associated with factorization. We demonstrate this method applied to the GPT-2$_{small}$. The result of the compression is TQCompressedGPT-2 model, featuring 81 mln. parameters compared to 124 mln. in the GPT-2$_{small}$. We make TQCompressedGPT-2 publicly available. We further enhance the performance of the TQCompressedGPT-2 through a training strategy involving multi-step knowledge distillation, using only a 3.1% of the OpenWebText. TQCompressedGPT-2 surpasses DistilGPT-2 and KnGPT-2 in comparative evaluations, marking an advancement in the efficient and effective deployment of models in resource-constrained environments.
Federated Learning (FL) is a promising technique for the collaborative training of deep neural networks across multiple devices while preserving data privacy. Despite its potential benefits, FL is hindered by excessive communication costs due to repeated server-client communication during training. To address this challenge, model compression techniques, such as sparsification and weight clustering are applied, which often require modifying the underlying model aggregation schemes or involve cumbersome hyperparameter tuning, with the latter not only adjusts the model's compression rate but also limits model's potential for continuous improvement over growing data. In this paper, we propose FedCompress, a novel approach that combines dynamic weight clustering and server-side knowledge distillation to reduce communication costs while learning highly generalizable models. Through a comprehensive evaluation on diverse public datasets, we demonstrate the efficacy of our approach compared to baselines in terms of communication costs and inference speed.
Deep clustering has gained significant attention due to its capability in learning clustering-friendly representations without labeled data. However, previous deep clustering methods tend to treat all samples equally, which neglect the variance in the latent distribution and the varying difficulty in classifying or clustering different samples. To address this, this paper proposes a novel end-to-end deep clustering method with diffused sampling and hardness-aware self-distillation (HaDis). Specifically, we first align one view of instances with another view via diffused sampling alignment (DSA), which helps improve the intra-cluster compactness. To alleviate the sampling bias, we present the hardness-aware self-distillation (HSD) mechanism to mine the hardest positive and negative samples and adaptively adjust their weights in a self-distillation fashion, which is able to deal with the potential imbalance in sample contributions during optimization. Further, the prototypical contrastive learning is incorporated to simultaneously enhance the inter-cluster separability and intra-cluster compactness. Experimental results on five challenging image datasets demonstrate the superior clustering performance of our HaDis method over the state-of-the-art. Source code is available at https://github.com/Regan-Zhang/HaDis.
Emotion Recognition in Conversation (ERC) plays a crucial role in enabling dialogue systems to effectively respond to user requests. The emotions in a conversation can be identified by the representations from various modalities, such as audio, visual, and text. However, due to the weak contribution of non-verbal modalities to recognize emotions, multimodal ERC has always been considered a challenging task. In this paper, we propose Teacher-leading Multimodal fusion network for ERC (TelME). TelME incorporates cross-modal knowledge distillation to transfer information from a language model acting as the teacher to the non-verbal students, thereby optimizing the efficacy of the weak modalities. We then combine multimodal features using a shifting fusion approach in which student networks support the teacher. TelME achieves state-of-the-art performance in MELD, a multi-speaker conversation dataset for ERC. Finally, we demonstrate the effectiveness of our components through additional experiments.
Natural Language Processing models like BERT can provide state-of-the-art word embeddings for downstream NLP tasks. However, these models yet to perform well on Semantic Textual Similarity, and may be too large to be deployed as lightweight edge applications. We seek to apply a suitable contrastive learning method based on the SimCSE paper, to a model architecture adapted from a knowledge distillation based model, DistilBERT, to address these two issues. Our final lightweight model DistilFace achieves an average of 72.1 in Spearman's correlation on STS tasks, a 34.2 percent improvement over BERT base.
Knowledge distillation has emerged as a highly effective method for bridging the representation discrepancy between large-scale models and lightweight models. Prevalent approaches involve leveraging appropriate metrics to minimize the divergence or distance between the knowledge extracted from the teacher model and the knowledge learned by the student model. Centered Kernel Alignment (CKA) is widely used to measure representation similarity and has been applied in several knowledge distillation methods. However, these methods are complex and fail to uncover the essence of CKA, thus not answering the question of how to use CKA to achieve simple and effective distillation properly. This paper first provides a theoretical perspective to illustrate the effectiveness of CKA, which decouples CKA to the upper bound of Maximum Mean Discrepancy~(MMD) and a constant term. Drawing from this, we propose a novel Relation-Centered Kernel Alignment~(RCKA) framework, which practically establishes a connection between CKA and MMD. Furthermore, we dynamically customize the application of CKA based on the characteristics of each task, with less computational source yet comparable performance than the previous methods. The extensive experiments on the CIFAR-100, ImageNet-1k, and MS-COCO demonstrate that our method achieves state-of-the-art performance on almost all teacher-student pairs for image classification and object detection, validating the effectiveness of our approaches. Our code is available in https://github.com/Klayand/PCKA
Advanced change detection techniques primarily target image pairs of equal and high quality. However, variations in imaging conditions and platforms frequently lead to image pairs with distinct qualities: one image being high-quality, while the other being low-quality. These disparities in image quality present significant challenges for understanding image pairs semantically and extracting change features, ultimately resulting in a notable decline in performance. To tackle this challenge, we introduce an innovative training strategy grounded in knowledge distillation. The core idea revolves around leveraging task knowledge acquired from high-quality image pairs to guide the model's learning process when dealing with image pairs that exhibit differences in quality. Additionally, we develop a hierarchical correlation distillation approach (involving self-correlation, cross-correlation, and global correlation). This approach compels the student model to replicate the correlations inherent in the teacher model, rather than focusing solely on individual features. This ensures effective knowledge transfer while maintaining the student model's training flexibility.
Recent advances in large language models (LLMs) have promoted generative error correction (GER) for automatic speech recognition (ASR), which leverages the rich linguistic knowledge and powerful reasoning ability of LLMs to improve recognition results. The latest work proposes a GER benchmark with HyPoradise dataset to learn the mapping from ASR N-best hypotheses to ground-truth transcription by efficient LLM finetuning, which shows great effectiveness but lacks specificity on noise-robust ASR. In this work, we extend the benchmark to noisy conditions and investigate if we can teach LLMs to perform denoising for GER just like what robust ASR do}, where one solution is introducing noise information as a conditioner into LLM. However, directly incorporating noise embeddings from audio encoder could harm the LLM tuning due to cross-modality gap. To this end, we propose to extract a language-space noise embedding from the N-best list to represent the noise conditions of source speech, which can promote the denoising process in GER. Furthermore, in order to enhance its representation ability of audio noise, we design a knowledge distillation (KD) approach via mutual information estimation to distill the real noise information in audio embeddings to our language embedding. Experiments on various latest LLMs demonstrate our approach achieves a new breakthrough with up to 53.9% correction improvement in terms of word error rate while with limited training data. Analysis shows that our language-space noise embedding can well represent the noise conditions of source speech, under which off-the-shelf LLMs show strong ability of language-space denoising.
The development of deep learning algorithms has extensively empowered humanity's task automatization capacity. However, the huge improvement in the performance of these models is highly correlated with their increasing level of complexity, limiting their usefulness in human-oriented applications, which are usually deployed in resource-constrained devices. This led to the development of compression techniques that drastically reduce the computational and memory costs of deep learning models without significant performance degradation. This paper aims to systematize the current literature on this topic by presenting a comprehensive survey of model compression techniques in biometrics applications, namely quantization, knowledge distillation and pruning. We conduct a critical analysis of the comparative value of these techniques, focusing on their advantages and disadvantages and presenting suggestions for future work directions that can potentially improve the current methods. Additionally, we discuss and analyze the link between model bias and model compression, highlighting the need to direct compression research toward model fairness in future works.
It is believed that in knowledge distillation (KD), the role of the teacher is to provide an estimate for the unknown Bayes conditional probability distribution (BCPD) to be used in the student training process. Conventionally, this estimate is obtained by training the teacher using maximum log-likelihood (MLL) method. To improve this estimate for KD, in this paper we introduce the concept of conditional mutual information (CMI) into the estimation of BCPD and propose a novel estimator called the maximum CMI (MCMI) method. Specifically, in MCMI estimation, both the log-likelihood and CMI of the teacher are simultaneously maximized when the teacher is trained. Through Eigen-CAM, it is further shown that maximizing the teacher's CMI value allows the teacher to capture more contextual information in an image cluster. Via conducting a thorough set of experiments, we show that by employing a teacher trained via MCMI estimation rather than one trained via MLL estimation in various state-of-the-art KD frameworks, the student's classification accuracy consistently increases, with the gain of up to 3.32\%. This suggests that the teacher's BCPD estimate provided by MCMI method is more accurate than that provided by MLL method. In addition, we show that such improvements in the student's accuracy are more drastic in zero-shot and few-shot settings. Notably, the student's accuracy increases with the gain of up to 5.72\% when 5\% of the training samples are available to the student (few-shot), and increases from 0\% to as high as 84\% for an omitted class (zero-shot). The code is available at \url{https://github.com/iclr2024mcmi/ICLRMCMI}.
Instance segmentation in remote sensing images is a long-standing challenge. Since horizontal bounding boxes introduce many interference objects, oriented bounding boxes (OBBs) are usually used for instance identification. However, based on ``segmentation within bounding box'' paradigm, current instance segmentation methods using OBBs are overly dependent on bounding box detection performance. To tackle this problem, this paper proposes OBSeg, an accurate and fast instance segmentation framework using OBBs. OBSeg is based on box prompt-based segmentation foundation models (BSMs), e.g., Segment Anything Model. Specifically, OBSeg first detects OBBs to distinguish instances and provide coarse localization information. Then, it predicts OBB prompt-related masks for fine segmentation. Since OBBs only serve as prompts, OBSeg alleviates the over-dependence on bounding box detection performance of current instance segmentation methods using OBBs. Thanks to OBB prompts, OBSeg outperforms other current BSM-based methods using HBBs. In addition, to enable BSMs to handle OBB prompts, we propose a novel OBB prompt encoder. To make OBSeg more lightweight and further improve the performance of lightweight distilled BSMs, a Gaussian smoothing-based knowledge distillation method is introduced. Experiments demonstrate that OBSeg outperforms current instance segmentation methods on multiple datasets in terms of instance segmentation accuracy and has competitive inference speed. The code is available at https://github.com/zhen6618/OBBInstanceSegmentation.
Knowledge distillation is the process of transferring knowledge from a more powerful large model (teacher) to a simpler counterpart (student). Numerous current approaches involve the student imitating the knowledge of the teacher directly. However, redundancy still exists in the learned representations through these prevalent methods, which tend to learn each spatial location's features indiscriminately. To derive a more compact representation (concept feature) from the teacher, inspired by human cognition, we suggest an innovative method, termed Generative Denoise Distillation (GDD), where stochastic noises are added to the concept feature of the student to embed them into the generated instance feature from a shallow network. Then, the generated instance feature is aligned with the knowledge of the instance from the teacher. We extensively experiment with object detection, instance segmentation, and semantic segmentation to demonstrate the versatility and effectiveness of our method. Notably, GDD achieves new state-of-the-art performance in the tasks mentioned above. We have achieved substantial improvements in semantic segmentation by enhancing PspNet and DeepLabV3, both of which are based on ResNet-18, resulting in mIoU scores of 74.67 and 77.69, respectively, surpassing their previous scores of 69.85 and 73.20 on the Cityscapes dataset of 20 categories. The source code is available at https://github.com/ZhgLiu/GDD.
Instance-aware embeddings predicted by deep neural networks have revolutionized biomedical instance segmentation, but its resource requirements are substantial. Knowledge distillation offers a solution by transferring distilled knowledge from heavy teacher networks to lightweight yet high-performance student networks. However, existing knowledge distillation methods struggle to extract knowledge for distinguishing instances and overlook global relation information. To address these challenges, we propose a graph relation distillation approach for efficient biomedical instance segmentation, which considers three essential types of knowledge: instance-level features, instance relations, and pixel-level boundaries. We introduce two graph distillation schemes deployed at both the intra-image level and the inter-image level: instance graph distillation (IGD) and affinity graph distillation (AGD). IGD constructs a graph representing instance features and relations, transferring these two types of knowledge by enforcing instance graph consistency. AGD constructs an affinity graph representing pixel relations to capture structured knowledge of instance boundaries, transferring boundary-related knowledge by ensuring pixel affinity consistency. Experimental results on a number of biomedical datasets validate the effectiveness of our approach, enabling student models with less than $ 1\%$ parameters and less than $10\%$ inference time while achieving promising performance compared to teacher models.