The advent of large language models (LLMs) revolutionized natural language processing applications, and running LLMs on edge devices has become increasingly attractive for reasons including reduced latency, data localization, and personalized user experiences. This comprehensive review examines the challenges of deploying computationally expensive LLMs on resource-constrained devices and explores innovative solutions across multiple domains. The paper investigates the development of on-device language models, their efficient architectures, including parameter sharing and modular designs, as well as state-of-the-art compression techniques like quantization, pruning, and knowledge distillation. Hardware acceleration strategies and collaborative edge-cloud deployment approaches are analyzed, highlighting the intricate balance between performance and resource utilization. Case studies of on-device language models from major mobile manufacturers demonstrate real-world applications and potential benefits. The review also addresses critical aspects such as adaptive learning, multi-modal capabilities, and personalization. By identifying key research directions and open challenges, this paper provides a roadmap for future advancements in on-device language models, emphasizing the need for interdisciplinary efforts to realize the full potential of ubiquitous, intelligent computing while ensuring responsible and ethical deployment. For a comprehensive review of research work and educational resources on on-device large language models (LLMs), please visit https://github.com/NexaAI/Awesome-LLMs-on-device. To download and run on-device LLMs, visit https://www.nexaai.com/models.
We present a novel frequency-based Self-Supervised Learning (SSL) approach that significantly enhances its efficacy for pre-training. Prior work in this direction masks out pre-defined frequencies in the input image and employs a reconstruction loss to pre-train the model. While achieving promising results, such an implementation has two fundamental limitations as identified in our paper. First, using pre-defined frequencies overlooks the variability of image frequency responses. Second, pre-trained with frequency-filtered images, the resulting model needs relatively more data to adapt to naturally looking images during fine-tuning. To address these drawbacks, we propose FOurier transform compression with seLf-Knowledge distillation (FOLK), integrating two dedicated ideas. First, inspired by image compression, we adaptively select the masked-out frequencies based on image frequency responses, creating more suitable SSL tasks for pre-training. Second, we employ a two-branch framework empowered by knowledge distillation, enabling the model to take both the filtered and original images as input, largely reducing the burden of downstream tasks. Our experimental results demonstrate the effectiveness of FOLK in achieving competitive performance to many state-of-the-art SSL methods across various downstream tasks, including image classification, few-shot learning, and semantic segmentation.
Autonomous driving holds great potential to transform road safety and traffic efficiency by minimizing human error and reducing congestion. A key challenge in realizing this potential is the accurate estimation of steering angles, which is essential for effective vehicle navigation and control. Recent breakthroughs in deep learning have made it possible to estimate steering angles directly from raw camera inputs. However, the limited available navigation data can hinder optimal feature learning, impacting the system's performance in complex driving scenarios. In this paper, we propose a shared encoder trained on multiple computer vision tasks critical for urban navigation, such as depth, pose, and 3D scene flow estimation, as well as semantic, instance, panoptic, and motion segmentation. By incorporating diverse visual information used by humans during navigation, this unified encoder might enhance steering angle estimation. To achieve effective multi-task learning within a single encoder, we introduce a multi-scale feature network for pose estimation to improve depth learning. Additionally, we employ knowledge distillation from a multi-backbone model pretrained on these navigation tasks to stabilize training and boost performance. Our findings demonstrate that a shared backbone trained on diverse visual tasks is capable of providing overall perception capabilities. While our performance in steering angle estimation is comparable to existing methods, the integration of human-like perception through multi-task learning holds significant potential for advancing autonomous driving systems. More details and the pretrained model are available at https://hi-computervision.github.io/uni-encoder/.
Speech restoration aims at restoring full-band speech with high quality and intelligibility, considering a diverse set of distortions. MaskSR is a recently proposed generative model for this task. As other models of its kind, MaskSR attains high quality but, as we show, intelligibility can be substantially improved. We do so by boosting the speech encoder component of MaskSR with predictions of semantic representations of the target speech, using a pre-trained self-supervised teacher model. Then, a masked language model is conditioned on the learned semantic features to predict acoustic tokens that encode low level spectral details of the target speech. We show that, with the same MaskSR model capacity and inference time, the proposed model, MaskSR2, significantly reduces the word error rate, a typical metric for intelligibility. MaskSR2 also achieves competitive word error rate among other models, while providing superior quality. An ablation study shows the effectiveness of various semantic representations.
Class Incremental Semantic Segmentation (CISS) aims to mitigate catastrophic forgetting by maintaining a balance between previously learned and newly introduced knowledge. Existing methods, primarily based on regularization techniques like knowledge distillation, help preserve old knowledge but often face challenges in effectively integrating new knowledge, resulting in limited overall improvement. Endpoints Weight Fusion (EWF) method, while simple, effectively addresses some of these limitations by dynamically fusing the model weights from previous steps with those from the current step, using a fusion parameter alpha determined by the relative number of previously known classes and newly introduced classes. However, the simplicity of the alpha calculation may limit its ability to fully capture the complexities of different task scenarios, potentially leading to suboptimal fusion outcomes. In this paper, we propose an enhanced approach called Adaptive Weight Fusion (AWF), which introduces an alternating training strategy for the fusion parameter, allowing for more flexible and adaptive weight integration. AWF achieves superior performance by better balancing the retention of old knowledge with the learning of new classes, significantly improving results on benchmark CISS tasks compared to the original EWF. And our experiment code will be released on Github.
Typically, the significant efficiency can be achieved by deploying different edge AI models in various real world scenarios while a few large models manage those edge AI models remotely from cloud servers. However, customizing edge AI models for each user's specific application or extending current models to new application scenarios remains a challenge. Inappropriate local training or fine tuning of edge AI models by users can lead to model malfunction, potentially resulting in legal issues for the manufacturer. To address aforementioned issues, this paper proposes an innovative framework called "DiReD", which involves knowledge DIstillation & REverse DIstillation. In the initial step, an edge AI model is trained with presumed data and a KD process using the cloud AI model in the upper management cloud server. This edge AI model is then dispatched to edge AI devices solely for inference in the user's application scenario. When the user needs to update the edge AI model to better fit the actual scenario, the reverse distillation (RD) process is employed to extract the knowledge: the difference between user preferences and the manufacturer's presumptions from the edge AI model using the user's exclusive data. Only the extracted knowledge is reported back to the upper management cloud server to update the cloud AI model, thus protecting user privacy by not using any exclusive data. The updated cloud AI can then update the edge AI model with the extended knowledge. Simulation results demonstrate that the proposed "DiReDi" framework allows the manufacturer to update the user model by learning new knowledge from the user's actual scenario with private data. The initial redundant knowledge is reduced since the retraining emphasizes user private data.
The Complex Emotion Recognition System (CERS) deciphers complex emotional states by examining combinations of basic emotions expressed, their interconnections, and the dynamic variations. Through the utilization of advanced algorithms, CERS provides profound insights into emotional dynamics, facilitating a nuanced understanding and customized responses. The attainment of such a level of emotional recognition in machines necessitates the knowledge distillation and the comprehension of novel concepts akin to human cognition. The development of AI systems for discerning complex emotions poses a substantial challenge with significant implications for affective computing. Furthermore, obtaining a sizable dataset for CERS proves to be a daunting task due to the intricacies involved in capturing subtle emotions, necessitating specialized methods for data collection and processing. Incorporating physiological signals such as Electrocardiogram (ECG) and Electroencephalogram (EEG) can notably enhance CERS by furnishing valuable insights into the user's emotional state, enhancing the quality of datasets, and fortifying system dependability. A comprehensive literature review was conducted in this study to assess the efficacy of machine learning, deep learning, and meta-learning approaches in both basic and complex emotion recognition utilizing EEG, ECG signals, and facial expression datasets. The chosen research papers offer perspectives on potential applications, clinical implications, and results of CERSs, with the objective of promoting their acceptance and integration into clinical decision-making processes. This study highlights research gaps and challenges in understanding CERSs, encouraging further investigation by relevant studies and organizations. Lastly, the significance of meta-learning approaches in improving CERS performance and guiding future research endeavors is underscored.
Cervical disc herniation (CDH) is a prevalent musculoskeletal disorder that significantly impacts health and requires labor-intensive analysis from experts. Despite advancements in automated detection of medical imaging, two significant challenges hinder the real-world application of these methods. First, the computational complexity and resource demands present a significant gap for real-time application. Second, noise in MRI reduces the effectiveness of existing methods by distorting feature extraction. To address these challenges, we propose three key contributions: Firstly, we introduced MedDet, which leverages the multi-teacher single-student knowledge distillation for model compression and efficiency, meanwhile integrating generative adversarial training to enhance performance. Additionally, we customize the second-order nmODE to improve the model's resistance to noise in MRI. Lastly, we conducted comprehensive experiments on the CDH-1848 dataset, achieving up to a 5% improvement in mAP compared to previous methods. Our approach also delivers over 5 times faster inference speed, with approximately 67.8% reduction in parameters and 36.9% reduction in FLOPs compared to the teacher model. These advancements significantly enhance the performance and efficiency of automated CDH detection, demonstrating promising potential for future application in clinical practice. See project website https://steve-zeyu-zhang.github.io/MedDet
Transformers have led to learning-based image compression methods that outperform traditional approaches. However, these methods often suffer from high complexity, limiting their practical application. To address this, various strategies such as knowledge distillation and lightweight architectures have been explored, aiming to enhance efficiency without significantly sacrificing performance. This paper proposes a State Space Model-based Image Compression (SSMIC) architecture. This novel architecture balances performance and computational efficiency, making it suitable for real-world applications. Experimental evaluations confirm the effectiveness of our model in achieving a superior BD-rate while significantly reducing computational complexity and latency compared to competitive learning-based image compression methods.
We report the development of Ruri, a series of Japanese general text embedding models. While the development of general-purpose text embedding models in English and multilingual contexts has been active in recent years, model development in Japanese remains insufficient. The primary reasons for this are the lack of datasets and the absence of necessary expertise. In this report, we provide a detailed account of the development process of Ruri. Specifically, we discuss the training of embedding models using synthesized datasets generated by LLMs, the construction of the reranker for dataset filtering and knowledge distillation, and the performance evaluation of the resulting general-purpose text embedding models.
Federated Learning (FL) is gaining popularity as a distributed learning framework that only shares model parameters or gradient updates and keeps private data locally. However, FL is at risk of privacy leakage caused by privacy inference attacks. And most existing privacy-preserving mechanisms in FL conflict with achieving high performance and efficiency. Therefore, we propose FedMD-CG, a novel FL method with highly competitive performance and high-level privacy preservation, which decouples each client's local model into a feature extractor and a classifier, and utilizes a conditional generator instead of the feature extractor to perform server-side model aggregation. To ensure the consistency of local generators and classifiers, FedMD-CG leverages knowledge distillation to train local models and generators at both the latent feature level and the logit level. Also, we construct additional classification losses and design new diversity losses to enhance client-side training. FedMD-CG is robust to data heterogeneity and does not require training extra discriminators (like cGAN). We conduct extensive experiments on various image classification tasks to validate the superiority of FedMD-CG.
Knowledge Distillation (KD) transfers knowledge from a large pre-trained teacher network to a compact and efficient student network, making it suitable for deployment on resource-limited media terminals. However, traditional KD methods require balanced data to ensure robust training, which is often unavailable in practical applications. In such scenarios, a few head categories occupy a substantial proportion of examples. This imbalance biases the trained teacher network towards the head categories, resulting in severe performance degradation on the less represented tail categories for both the teacher and student networks. In this paper, we propose a novel framework called Knowledge Rectification Distillation (KRDistill) to address the imbalanced knowledge inherited in the teacher network through the incorporation of the balanced category priors. Furthermore, we rectify the biased predictions produced by the teacher network, particularly focusing on the tail categories. Consequently, the teacher network can provide balanced and accurate knowledge to train a reliable student network. Intensive experiments conducted on various long-tailed datasets demonstrate that our KRDistill can effectively train reliable student networks in realistic scenarios of data imbalance.
The application of machine learning to medical ultrasound videos of the heart, i.e., echocardiography, has recently gained traction with the availability of large public datasets. Traditional supervised tasks, such as ejection fraction regression, are now making way for approaches focusing more on the latent structure of data distributions, as well as generative methods. We propose a model trained exclusively by knowledge distillation, either on real or synthetical data, involving retrieving masks suggested by a teacher model. We achieve state-of-the-art (SOTA) values on the task of identifying end-diastolic and end-systolic frames. By training the model only on synthetic data, it reaches segmentation capabilities close to the performance when trained on real data with a significantly reduced number of weights. A comparison with the 5 main existing methods shows that our method outperforms the others in most cases. We also present a new evaluation method that does not require human annotation and instead relies on a large auxiliary model. We show that this method produces scores consistent with those obtained from human annotations. Relying on the integrated knowledge from a vast amount of records, this method overcomes certain inherent limitations of human annotator labeling. Code: https://github.com/GregoirePetit/EchoDFKD
In the field of Alzheimer's disease diagnosis, segmentation and classification tasks are inherently interconnected. Sharing knowledge between models for these tasks can significantly improve training efficiency, particularly when training data is scarce. However, traditional knowledge distillation techniques often struggle to bridge the gap between segmentation and classification due to the distinct nature of tasks and different model architectures. To address this challenge, we propose a dual-stream pipeline that facilitates cross-task and cross-architecture knowledge sharing. Our approach introduces a dual-stream embedding module that unifies feature representations from segmentation and classification models, enabling dimensional integration of these features to guide the classification model. We validated our method on multiple 3D datasets for Alzheimer's disease diagnosis, demonstrating significant improvements in classification performance, especially on small datasets. Furthermore, we extended our pipeline with a residual temporal attention mechanism for early diagnosis, utilizing images taken before the atrophy of patients' brain mass. This advancement shows promise in enabling diagnosis approximately six months earlier in mild and asymptomatic stages, offering critical time for intervention.
This paper presents LiteVSR2, an enhanced version of our previously introduced efficient approach to Visual Speech Recognition (VSR). Building upon our knowledge distillation framework from a pre-trained Automatic Speech Recognition (ASR) model, we introduce two key improvements: a stabilized video preprocessing technique and feature normalization in the distillation process. These improvements yield substantial performance gains on the LRS2 and LRS3 benchmarks, positioning LiteVSR2 as the current best CTC-based VSR model without increasing the volume of training data or computational resources utilized. Furthermore, we explore the scalability of our approach by examining performance metrics across varying model complexities and training data volumes. LiteVSR2 maintains the efficiency of its predecessor while significantly enhancing accuracy, thereby demonstrating the potential for resource-efficient advancements in VSR technology.
A new algorithm for incremental learning in the context of Tiny Machine learning (TinyML) is presented, which is optimized for low-performance and energy efficient embedded devices. TinyML is an emerging field that deploys machine learning models on resource-constrained devices such as microcontrollers, enabling intelligent applications like voice recognition, anomaly detection, predictive maintenance, and sensor data processing in environments where traditional machine learning models are not feasible. The algorithm solve the challenge of catastrophic forgetting through the use of knowledge distillation to create a small, distilled dataset. The novelty of the method is that the size of the model can be adjusted dynamically, so that the complexity of the model can be adapted to the requirements of the task. This offers a solution for incremental learning in resource-constrained environments, where both model size and computational efficiency are critical factors. Results show that the proposed algorithm offers a promising approach for TinyML incremental learning on embedded devices. The algorithm was tested on five datasets including: CIFAR10, MNIST, CORE50, HAR, Speech Commands. The findings indicated that, despite using only 43% of Floating Point Operations (FLOPs) compared to a larger fixed model, the algorithm experienced a negligible accuracy loss of just 1%. In addition, the presented method is memory efficient. While state-of-the-art incremental learning is usually very memory intensive, the method requires only 1% of the original data set.
The time-consuming nature of training and deploying complicated Machine and Deep Learning (DL) models for a variety of applications continues to pose significant challenges in the field of Machine Learning (ML). These challenges are particularly pronounced in the federated domain, where optimizing models for individual nodes poses significant difficulty. Many methods have been developed to tackle this problem, aiming to reduce training expenses and time while maintaining efficient optimisation. Three suggested strategies to tackle this challenge include Active Learning, Knowledge Distillation, and Local Memorization. These methods enable the adoption of smaller models that require fewer computational resources and allow for model personalization with local insights, thereby improving the effectiveness of current models. The present study delves into the fundamental principles of these three approaches and proposes an advanced Federated Learning System that utilises different Personalisation methods towards improving the accuracy of AI models and enhancing user experience in real-time NG-IoT applications, investigating the efficacy of these techniques in the local and federated domain. The results of the original and optimised models are then compared in both local and federated contexts using a comparison analysis. The post-analysis shows encouraging outcomes when it comes to optimising and personalising the models with the suggested techniques.
Spatio-temporal prediction is a crucial research area in data-driven urban computing, with implications for transportation, public safety, and environmental monitoring. However, scalability and generalization challenges remain significant obstacles. Advanced models often rely on Graph Neural Networks to encode spatial and temporal correlations, but struggle with the increased complexity of large-scale datasets. The recursive GNN-based message passing schemes used in these models hinder their training and deployment in real-life urban sensing scenarios. Moreover, long-spanning large-scale spatio-temporal data introduce distribution shifts, necessitating improved generalization performance. To address these challenges, we propose a simple framework for spatio-temporal prediction - EasyST paradigm. It learns lightweight and robust Multi-Layer Perceptrons (MLPs) by effectively distilling knowledge from complex spatio-temporal GNNs. We ensure robust knowledge distillation by integrating the spatio-temporal information bottleneck with teacher-bounded regression loss, filtering out task-irrelevant noise and avoiding erroneous guidance. We further enhance the generalization ability of the student model by incorporating spatial and temporal prompts to provide downstream task contexts. Evaluation on three spatio-temporal datasets for urban computing tasks demonstrates that EasyST surpasses state-of-the-art approaches in terms of efficiency and accuracy. The implementation code is available at: https://github.com/HKUDS/EasyST.
Transformers have revolutionized the object detection landscape by introducing DETRs, acclaimed for their simplicity and efficacy. Despite their advantages, the substantial size of these models poses significant challenges for practical deployment, particularly in resource-constrained environments. This paper addresses the challenge of compressing DETR by leveraging knowledge distillation, a technique that holds promise for maintaining model performance while reducing size. A critical aspect of DETRs' performance is their reliance on queries to interpret object representations accurately. Traditional distillation methods often focus exclusively on positive queries, identified through bipartite matching, neglecting the rich information present in hard-negative queries. Our visual analysis indicates that hard-negative queries, focusing on foreground elements, are crucial for enhancing distillation outcomes. To this end, we introduce a novel Group Query Selection strategy, which diverges from traditional query selection in DETR distillation by segmenting queries based on their Generalized Intersection over Union (GIoU) with ground truth objects, thereby uncovering valuable hard-negative queries for distillation. Furthermore, we present the Knowledge Distillation via Query Selection for DETR (QSKD) framework, which incorporates Attention-Guided Feature Distillation (AGFD) and Local Alignment Prediction Distillation (LAPD). These components optimize the distillation process by focusing on the most informative aspects of the teacher model's intermediate features and output. Our comprehensive experimental evaluation of the MS-COCO dataset demonstrates the effectiveness of our approach, significantly improving average precision (AP) across various DETR architectures without incurring substantial computational costs. Specifically, the AP of Conditional DETR ResNet-18 increased from 35.8 to 39.9.
Very low-resolution face recognition is challenging due to the serious loss of informative facial details in resolution degradation. In this paper, we propose a generative-discriminative representation distillation approach that combines generative representation with cross-resolution aligned knowledge distillation. This approach facilitates very low-resolution face recognition by jointly distilling generative and discriminative models via two distillation modules. Firstly, the generative representation distillation takes the encoder of a diffusion model pretrained for face super-resolution as the generative teacher to supervise the learning of the student backbone via feature regression, and then freezes the student backbone. After that, the discriminative representation distillation further considers a pretrained face recognizer as the discriminative teacher to supervise the learning of the student head via cross-resolution relational contrastive distillation. In this way, the general backbone representation can be transformed into discriminative head representation, leading to a robust and discriminative student model for very low-resolution face recognition. Our approach improves the recovery of the missing details in very low-resolution faces and achieves better knowledge transfer. Extensive experiments on face datasets demonstrate that our approach enhances the recognition accuracy of very low-resolution faces, showcasing its effectiveness and adaptability.