Federated Learning (FL) empowers multiple clients to collaboratively train machine learning models without sharing local data, making it highly applicable in heterogeneous Internet of Things (IoT) environments. However, intrinsic heterogeneity in clients' model architectures and computing capabilities often results in model accuracy loss and the intractable straggler problem, which significantly impairs training effectiveness. To tackle these challenges, this paper proposes a novel Heterogeneity-aware Personalized Federated Learning method, named HAPFL, via multi-level Reinforcement Learning (RL) mechanisms. HAPFL optimizes the training process by incorporating three strategic components: 1) An RL-based heterogeneous model allocation mechanism. The parameter server employs a Proximal Policy Optimization (PPO)-based RL agent to adaptively allocate appropriately sized, differentiated models to clients based on their performance, effectively mitigating performance disparities. 2) An RL-based training intensity adjustment scheme. The parameter server leverages another PPO-based RL agent to dynamically fine-tune the training intensity for each client to further enhance training efficiency and reduce straggling latency. 3) A knowledge distillation-based mutual learning mechanism. Each client deploys both a heterogeneous local model and a homogeneous lightweight model named LiteModel, where these models undergo mutual learning through knowledge distillation. This uniform LiteModel plays a pivotal role in aggregating and sharing global knowledge, significantly enhancing the effectiveness of personalized local training. Experimental results across multiple benchmark datasets demonstrate that HAPFL not only achieves high accuracy but also substantially reduces the overall training time by 20.9%-40.4% and decreases straggling latency by 19.0%-48.0% compared to existing solutions.
Causal language models have demonstrated remarkable capabilities, but their size poses significant challenges for deployment in resource-constrained environments. Knowledge distillation, a widely-used technique for transferring knowledge from a large teacher model to a small student model, presents a promising approach for model compression. A significant remaining issue lies in the major differences between teacher and student models, namely the substantial capacity gap, mode averaging, and mode collapse, which pose barriers during distillation. To address these issues, we introduce $\textit{Temporally Adaptive Interpolated Distillation (TAID)}$, a novel knowledge distillation approach that dynamically interpolates student and teacher distributions through an adaptive intermediate distribution, gradually shifting from the student's initial distribution towards the teacher's distribution. We provide a theoretical analysis demonstrating TAID's ability to prevent mode collapse and empirically show its effectiveness in addressing the capacity gap while balancing mode averaging and mode collapse. Our comprehensive experiments demonstrate TAID's superior performance across various model sizes and architectures in both instruction tuning and pre-training scenarios. Furthermore, we showcase TAID's practical impact by developing two state-of-the-art compact foundation models: $\texttt{TAID-LLM-1.5B}$ for language tasks and $\texttt{TAID-VLM-2B}$ for vision-language tasks. These results demonstrate TAID's effectiveness in creating high-performing and efficient models, advancing the development of more accessible AI technologies.
The Segment Anything Model (SAM) has set a new standard in interactive image segmentation, offering robust performance across various tasks. However, its significant computational requirements limit its deployment in real-time or resource-constrained environments. To address these challenges, we propose a novel knowledge distillation approach, KD SAM, which incorporates both encoder and decoder optimization through a combination of Mean Squared Error (MSE) and Perceptual Loss. This dual-loss framework captures structural and semantic features, enabling the student model to maintain high segmentation accuracy while reducing computational complexity. Based on the model evaluation on datasets, including Kvasir-SEG, ISIC 2017, Fetal Head Ultrasound, and Breast Ultrasound, we demonstrate that KD SAM achieves comparable or superior performance to the baseline models, with significantly fewer parameters. KD SAM effectively balances segmentation accuracy and computational efficiency, making it well-suited for real-time medical image segmentation applications in resource-constrained environments.
Spiking Neural Networks (SNNs) are emerging as a brain-inspired alternative to traditional Artificial Neural Networks (ANNs), prized for their potential energy efficiency on neuromorphic hardware. Despite this, SNNs often suffer from accuracy degradation compared to ANNs and face deployment challenges due to fixed inference timesteps, which require retraining for adjustments, limiting operational flexibility. To address these issues, our work considers the spatio-temporal property inherent in SNNs, and proposes a novel distillation framework for deep SNNs that optimizes performance across full-range timesteps without specific retraining, enhancing both efficacy and deployment adaptability. We provide both theoretical analysis and empirical validations to illustrate that training guarantees the convergence of all implicit models across full-range timesteps. Experimental results on CIFAR-10, CIFAR-100, CIFAR10-DVS, and ImageNet demonstrate state-of-the-art performance among distillation-based SNNs training methods. Our code is available at https://github.com/Intelli-Chip-Lab/snn\_temporal\_decoupling\_distillation.
Gait recognition is an important biometric technique over large distances. State-of-the-art gait recognition systems perform very well in controlled environments at close range. Recently, there has been an increased interest in gait recognition in the wild prompted by the collection of outdoor, more challenging datasets containing variations in terms of illumination, pitch angles, and distances. An important problem in these environments is that of occlusion, where the subject is partially blocked from camera view. While important, this problem has received little attention. Thus, we propose MimicGait, a model-agnostic approach for gait recognition in the presence of occlusions. We train the network using a multi-instance correlational distillation loss to capture both inter-sequence and intra-sequence correlations in the occluded gait patterns of a subject, utilizing an auxiliary Visibility Estimation Network to guide the training of the proposed mimic network. We demonstrate the effectiveness of our approach on challenging real-world datasets like GREW, Gait3D and BRIAR. We release the code in https://github.com/Ayush-00/mimicgait.
This paper aims to improve the robustness of a small global model while maintaining clean accuracy under adversarial attacks and non-IID challenges in federated learning. By leveraging the concise knowledge embedded in the class probabilities from a pre-trained model for both clean and adversarial image classification, we propose a Pre-trained Model-guided Adversarial Federated Learning (PM-AFL) training paradigm. This paradigm integrates vanilla mixture and adversarial mixture knowledge distillation to effectively balance accuracy and robustness while promoting local models to learn from diverse data. Specifically, for clean accuracy, we adopt a dual distillation strategy where the class probabilities of randomly paired images and their blended versions are aligned between the teacher model and the local models. For adversarial robustness, we use a similar distillation approach but replace clean samples on the local side with adversarial examples. Moreover, considering the bias between local and global models, we also incorporate a consistency regularization term to ensure that local adversarial predictions stay aligned with their corresponding global clean ones. These strategies collectively enable local models to absorb diverse knowledge from the teacher model while maintaining close alignment with the global model, thereby mitigating overfitting to local optima and enhancing the generalization of the global model. Experiments demonstrate that the PM-AFL-based paradigm outperforms other methods that integrate defense strategies by a notable margin.
Large language models (LLMs) have demonstrated immense capabilities in understanding textual data and are increasingly being adopted to help researchers accelerate scientific discovery through knowledge extraction (information retrieval), knowledge distillation (summarizing key findings and methodologies into concise forms), and knowledge synthesis (aggregating information from multiple scientific sources to address complex queries, generate hypothesis and formulate experimental plans). However, scientific data often exists in both visual and textual modalities. Vision language models (VLMs) address this by incorporating a pretrained vision backbone for processing images and a cross-modal projector that adapts image tokens into the LLM dimensional space, thereby providing richer multimodal comprehension. Nevertheless, off-the-shelf VLMs show limited capabilities in handling domain-specific data and are prone to hallucinations. We developed intelligent assistants finetuned from LLaVA models to enhance multimodal understanding in low-dose radiation therapy (LDRT)-a benign approach used in the treatment of cancer-related illnesses. Using multilingual data from 42,673 articles, we devise complex reasoning and detailed description tasks for visual question answering (VQA) benchmarks. Our assistants, trained on 50,882 image-text pairs, demonstrate superior performance over base models as evaluated using LLM-as-a-judge approach, particularly in reducing hallucination and improving domain-specific comprehension.
Video surveillance systems are crucial components for ensuring public safety and management in smart city. As a fundamental task in video surveillance, text-to-image person retrieval aims to retrieve the target person from an image gallery that best matches the given text description. Most existing text-to-image person retrieval methods are trained in a supervised manner that requires sufficient labeled data in the target domain. However, it is common in practice that only unlabeled data is available in the target domain due to the difficulty and cost of data annotation, which limits the generalization of existing methods in practical application scenarios. To address this issue, we propose a novel unsupervised domain adaptation method, termed Graph-Based Cross-Domain Knowledge Distillation (GCKD), to learn the cross-modal feature representation for text-to-image person retrieval in a cross-dataset scenario. The proposed GCKD method consists of two main components. Firstly, a graph-based multi-modal propagation module is designed to bridge the cross-domain correlation among the visual and textual samples. Secondly, a contrastive momentum knowledge distillation module is proposed to learn the cross-modal feature representation using the online knowledge distillation strategy. By jointly optimizing the two modules, the proposed method is able to achieve efficient performance for cross-dataset text-to-image person retrieval. acExtensive experiments on three publicly available text-to-image person retrieval datasets demonstrate the effectiveness of the proposed GCKD method, which consistently outperforms the state-of-the-art baselines.
Resource-constrained edge deployments demand AI solutions that balance high performance with stringent compute, memory, and energy limitations. In this survey, we present a comprehensive overview of the primary strategies for accelerating deep learning models under such constraints. First, we examine model compression techniques-pruning, quantization, tensor decomposition, and knowledge distillation-that streamline large models into smaller, faster, and more efficient variants. Next, we explore Neural Architecture Search (NAS), a class of automated methods that discover architectures inherently optimized for particular tasks and hardware budgets. We then discuss compiler and deployment frameworks, such as TVM, TensorRT, and OpenVINO, which provide hardware-tailored optimizations at inference time. By integrating these three pillars into unified pipelines, practitioners can achieve multi-objective goals, including latency reduction, memory savings, and energy efficiency-all while maintaining competitive accuracy. We also highlight emerging frontiers in hierarchical NAS, neurosymbolic approaches, and advanced distillation tailored to large language models, underscoring open challenges like pre-training pruning for massive networks. Our survey offers practical insights, identifies current research gaps, and outlines promising directions for building scalable, platform-independent frameworks to accelerate deep learning models at the edge.
The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.
Retrieval-Augmented Generation (RAG) pipelines enhance Large Language Models (LLMs) by retrieving relevant documents, but they face scalability issues due to high inference costs and limited context size. Document compression is a practical solution, but current soft compression methods suffer from accuracy losses and require extensive pretraining. In this paper, we introduce PISCO, a novel method that achieves a 16x compression rate with minimal accuracy loss (0-3%) across diverse RAG-based question-answering (QA) tasks. Unlike existing approaches, PISCO requires no pretraining or annotated data, relying solely on sequence-level knowledge distillation from document-based questions. With the ability to fine-tune a 7-10B LLM in 48 hours on a single A100 GPU, PISCO offers a highly efficient and scalable solution. We present comprehensive experiments showing that PISCO outperforms existing compression models by 8% in accuracy.
Dense retrievers have demonstrated significant potential for neural information retrieval; however, they exhibit a lack of robustness to domain shifts, thereby limiting their efficacy in zero-shot settings across diverse domains. A state-of-the-art domain adaptation technique is Generative Pseudo Labeling (GPL). GPL uses synthetic query generation and initially mined hard negatives to distill knowledge from cross-encoder to dense retrievers in the target domain. In this paper, we analyze the documents retrieved by the domain-adapted model and discover that these are more relevant to the target queries than those of the non-domain-adapted model. We then propose refreshing the hard-negative index during the knowledge distillation phase to mine better hard negatives. Our remining R-GPL approach boosts ranking performance in 13/14 BEIR datasets and 9/12 LoTTe datasets. Our contributions are (i) analyzing hard negatives returned by domain-adapted and non-domain-adapted models and (ii) applying the GPL training with and without hard-negative re-mining in LoTTE and BEIR datasets.
We introduce a multimodal deep learning framework, Prescriptive Neural Networks (PNNs), that combines ideas from optimization and machine learning, and is, to the best of our knowledge, the first prescriptive method to handle multimodal data. The PNN is a feedforward neural network trained on embeddings to output an outcome-optimizing prescription. In two real-world multimodal datasets, we demonstrate that PNNs prescribe treatments that are able to significantly improve estimated outcomes in transcatheter aortic valve replacement (TAVR) procedures by reducing estimated postoperative complication rates by 32% and in liver trauma injuries by reducing estimated mortality rates by over 40%. In four real-world, unimodal tabular datasets, we demonstrate that PNNs outperform or perform comparably to other well-known, state-of-the-art prescriptive models; importantly, on tabular datasets, we also recover interpretability through knowledge distillation, fitting interpretable Optimal Classification Tree models onto the PNN prescriptions as classification targets, which is critical for many real-world applications. Finally, we demonstrate that our multimodal PNN models achieve stability across randomized data splits comparable to other prescriptive methods and produce realistic prescriptions across the different datasets.
In computational optical imaging and wireless communications, signals are acquired through linear coded and noisy projections, which are recovered through computational algorithms. Deep model-based approaches, i.e., neural networks incorporating the sensing operators, are the state-of-the-art for signal recovery. However, these methods require exact knowledge of the sensing operator, which is often unavailable in practice, leading to performance degradation. Consequently, we propose a new recovery paradigm based on knowledge distillation. A teacher model, trained with full or almost exact knowledge of a synthetic sensing operator, guides a student model with an inexact real sensing operator. The teacher is interpreted as a relaxation of the student since it solves a problem with fewer constraints, which can guide the student to achieve higher performance. We demonstrate the improvement of signal reconstruction in computational optical imaging for single-pixel imaging with miscalibrated coded apertures systems and multiple-input multiple-output symbols detection with inexact channel matrix.
Graph neural networks (GNNs) have gained considerable attention in recent years for traffic flow prediction due to their ability to learn spatio-temporal pattern representations through a graph-based message-passing framework. Although GNNs have shown great promise in handling traffic datasets, their deployment in real-life applications has been hindered by scalability constraints arising from high-order message passing. Additionally, the over-smoothing problem of GNNs may lead to indistinguishable region representations as the number of layers increases, resulting in performance degradation. To address these challenges, we propose a new knowledge distillation paradigm termed LightST that transfers spatial and temporal knowledge from a high-capacity teacher to a lightweight student. Specifically, we introduce a spatio-temporal knowledge distillation framework that helps student MLPs capture graph-structured global spatio-temporal patterns while alleviating the over-smoothing effect with adaptive knowledge distillation. Extensive experiments verify that LightST significantly speeds up traffic flow predictions by 5X to 40X compared to state-of-the-art spatio-temporal GNNs, all while maintaining superior accuracy.
Recent advancements in deep learning have significantly improved performance on computer vision tasks. Previous image classification methods primarily modify model architectures or add features, and they optimize models using cross-entropy loss on class logits. Since they focus on classifying images with considering class labels, these methods may struggle to learn various \emph{aspects} of classes (e.g., natural positions and shape changes). Rethinking the previous approach from a novel view, we propose a multi-aspect knowledge distillation method using Multimodal Large Language Models (MLLMs). Our approach involves: 1) querying Large Language Model with multi-aspect questions relevant to the knowledge we want to transfer to the model, 2) extracting corresponding logits from MLLM, and 3) expanding the model's output dimensions to distill these multi-aspect logits. We then apply cross-entropy loss to class logits and binary cross-entropy loss to multi-aspect logits. Through our method, the model can learn not only the knowledge about visual aspects but also the abstract and complex aspects that require a deeper understanding. We primarily apply our method to image classification, and to explore the potential for extending our model, we expand it to other tasks, such as object detection. In all experimental results, our method improves the performance of the baselines. Additionally, we analyze the effect of multi-aspect knowledge distillation. These results demonstrate that our method can transfer knowledge about various aspects to the model and the aspect knowledge can enhance model performance in computer vision tasks. This paper demonstrates the great potential of multi-aspect knowledge distillation, and we believe it offers a promising direction for future research in computer vision and beyond.
Federated Learning (FL) has emerged as a prominent distributed learning paradigm. Within the scope of privacy preservation, information privacy regulations such as GDPR entitle users to request the removal (or unlearning) of their contribution from a service that is hosting the model. For this purpose, a server hosting an ML model must be able to unlearn certain information in cases such as copyright infringement or security issues that can make the model vulnerable or impact the performance of a service based on that model. While most unlearning approaches in FL focus on Horizontal FL (HFL), where clients share the feature space and the global model, Vertical FL (VFL) has received less attention from the research community. VFL involves clients (passive parties) sharing the sample space among them while not having access to the labels. In this paper, we explore unlearning in VFL from three perspectives: unlearning clients, unlearning features, and unlearning samples. To unlearn clients and features we introduce VFU-KD which is based on knowledge distillation (KD) while to unlearn samples, VFU-GA is introduced which is based on gradient ascent. To provide evidence of approximate unlearning, we utilize Membership Inference Attack (MIA) to audit the effectiveness of our unlearning approach. Our experiments across six tabular datasets and two image datasets demonstrate that VFU-KD and VFU-GA achieve performance comparable to or better than both retraining from scratch and the benchmark R2S method in many cases, with improvements of $(0-2\%)$. In the remaining cases, utility scores remain comparable, with a modest utility loss ranging from $1-5\%$. Unlike existing methods, VFU-KD and VFU-GA require no communication between active and passive parties during unlearning. However, they do require the active party to store the previously communicated embeddings.
In commonly used sub-quadratic complexity modules, linear attention benefits from simplicity and high parallelism, making it promising for image synthesis tasks. However, the architectural design and learning strategy for linear attention remain underexplored in this field. In this paper, we offer a suite of ready-to-use solutions for efficient linear diffusion Transformers. Our core contributions include: (1) Simplified Linear Attention using few heads, observing the free-lunch effect of performance without latency increase. (2) Weight inheritance from a fully pre-trained diffusion Transformer: initializing linear Transformer using pre-trained diffusion Transformer and loading all parameters except for those related to linear attention. (3) Hybrid knowledge distillation objective: using a pre-trained diffusion Transformer to help the training of the student linear Transformer, supervising not only the predicted noise but also the variance of the reverse diffusion process. These guidelines lead to our proposed Linear Diffusion Transformer (LiT), an efficient text-to-image Transformer that can be deployed offline on a laptop. Experiments show that in class-conditional 256*256 and 512*512 ImageNet benchmark LiT achieves highly competitive FID while reducing training steps by 80% and 77% compared to DiT. LiT also rivals methods based on Mamba or Gated Linear Attention. Besides, for text-to-image generation, LiT allows for the rapid synthesis of up to 1K resolution photorealistic images. Project page: https://techmonsterwang.github.io/LiT/.
Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 60% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge sharing among requests. However, naively caching and reusing past responses leads to large quality degradation. In this paper, we introduce EchoLM, an in-context caching system that leverages historical requests as examples to guide response generation, enabling selective offloading of requests to more efficient LLMs. However, enabling this real-time knowledge transfer leads to intricate tradeoffs between response quality, latency, and system throughput at scale. For a new request, EchoLM identifies similar, high-utility examples and efficiently prepends them to the input for better response. At scale, EchoLM adaptively routes requests to LLMs of varying capabilities, accounting for response quality and serving loads. EchoLM employs a cost-aware cache replay mechanism to improve example quality and coverage offline, maximizing cache utility and runtime efficiency. Evaluations on millions of open-source requests demonstrate that EchoLM has a throughput improvement of 1.4-5.9x while reducing latency by 28-71% without hurting response quality on average.
In this paper, we propose the use of simple knowledge distillation to produce smaller and more efficient single-language transformers from Massively Multilingual Transformers (MMTs) to alleviate tradeoffs associated with the use of such in low-resource settings. Using Tagalog as a case study, we show that these smaller single-language models perform on-par with strong baselines in a variety of benchmark tasks in a much more efficient manner. Furthermore, we investigate additional steps during the distillation process that improves the soft-supervision of the target language, and provide a number of analyses and ablations to show the efficacy of the proposed method.