Knowledge distillation (KD) has emerged as a promising technique in deep learning, typically employed to enhance a compact student network through learning from their high-performance but more complex teacher variant. When applied in the context of image super-resolution, most KD approaches are modified versions of methods developed for other computer vision tasks, which are based on training strategies with a single teacher and simple loss functions. In this paper, we propose a novel Multi-Teacher Knowledge Distillation (MTKD) framework specifically for image super-resolution. It exploits the advantages of multiple teachers by combining and enhancing the outputs of these teacher models, which then guides the learning process of the compact student network. To achieve more effective learning performance, we have also developed a new wavelet-based loss function for MTKD, which can better optimize the training process by observing differences in both the spatial and frequency domains. We fully evaluate the effectiveness of the proposed method by comparing it to five commonly used KD methods for image super-resolution based on three popular network architectures. The results show that the proposed MTKD method achieves evident improvements in super-resolution performance, up to 0.46dB (based on PSNR), over state-of-the-art KD approaches across different network structures. The source code of MTKD will be made available here for public evaluation.
Face Image Quality Assessment (FIQA) techniques have seen steady improvements over recent years, but their performance still deteriorates if the input face samples are not properly aligned. This alignment sensitivity comes from the fact that most FIQA techniques are trained or designed using a specific face alignment procedure. If the alignment technique changes, the performance of most existing FIQA techniques quickly becomes suboptimal. To address this problem, we present in this paper a novel knowledge distillation approach, termed AI-KD that can extend on any existing FIQA technique, improving its robustness to alignment variations and, in turn, performance with different alignment procedures. To validate the proposed distillation approach, we conduct comprehensive experiments on 6 face datasets with 4 recent face recognition models and in comparison to 7 state-of-the-art FIQA techniques. Our results show that AI-KD consistently improves performance of the initial FIQA techniques not only with misaligned samples, but also with properly aligned facial images. Furthermore, it leads to a new state-of-the-art, when used with a competitive initial FIQA approach. The code for AI-KD is made publicly available from: https://github.com/LSIbabnikz/AI-KD.
Few-shot knowledge distillation recently emerged as a viable approach to harness the knowledge of large-scale pre-trained models, using limited data and computational resources. In this paper, we propose a novel few-shot feature distillation approach for vision transformers. Our approach is based on two key steps. Leveraging the fact that vision transformers have a consistent depth-wise structure, we first copy the weights from intermittent layers of existing pre-trained vision transformers (teachers) into shallower architectures (students), where the intermittence factor controls the complexity of the student transformer with respect to its teacher. Next, we employ an enhanced version of Low-Rank Adaptation (LoRA) to distill knowledge into the student in a few-shot scenario, aiming to recover the information processing carried out by the skipped teacher layers. We present comprehensive experiments with supervised and self-supervised transformers as teachers, on six data sets from various domains (natural, medical and satellite images) and tasks (classification and segmentation). The empirical results confirm the superiority of our approach over state-of-the-art competitors. Moreover, the ablation results demonstrate the usefulness of each component of the proposed pipeline. We release our code at https://github.com/dianagrigore/WeCoLoRA.
With the surge of ChatGPT,the use of large models has significantly increased,rapidly rising to prominence across the industry and sweeping across the internet. This article is a comprehensive review of fine-tuning methods for large models. This paper investigates the latest technological advancements and the application of advanced methods in aspects such as task-adaptive fine-tuning,domain-adaptive fine-tuning,few-shot learning,knowledge distillation,multi-task learning,parameter-efficient fine-tuning,and dynamic fine-tuning.
Federated Learning (FL) seeks to train a model collaboratively without sharing private training data from individual clients. Despite its promise, FL encounters challenges such as high communication costs for large-scale models and the necessity for uniform model architectures across all clients and the server. These challenges severely restrict the practical applications of FL. To address these limitations, the integration of knowledge distillation (KD) into FL has been proposed, forming what is known as Federated Distillation (FD). FD enables more flexible knowledge transfer between clients and the server, surpassing the mere sharing of model parameters. By eliminating the need for identical model architectures across clients and the server, FD mitigates the communication costs associated with training large-scale models. This paper aims to offer a comprehensive overview of FD, highlighting its latest advancements. It delves into the fundamental principles underlying the design of FD frameworks, delineates FD approaches for tackling various challenges, and provides insights into the diverse applications of FD across different scenarios.
Inferring scene geometry from images via Structure from Motion is a long-standing and fundamental problem in computer vision. While classical approaches and, more recently, depth map predictions only focus on the visible parts of a scene, the task of scene completion aims to reason about geometry even in occluded regions. With the popularity of neural radiance fields (NeRFs), implicit representations also became popular for scene completion by predicting so-called density fields. Unlike explicit approaches. e.g. voxel-based methods, density fields also allow for accurate depth prediction and novel-view synthesis via image-based rendering. In this work, we propose to fuse the scene reconstruction from multiple images and distill this knowledge into a more accurate single-view scene reconstruction. To this end, we propose Multi-View Behind the Scenes (MVBTS) to fuse density fields from multiple posed images, trained fully self-supervised only from image data. Using knowledge distillation, we use MVBTS to train a single-view scene completion network via direct supervision called KDBTS. It achieves state-of-the-art performance on occupancy prediction, especially in occluded regions.
Blind-spot networks (BSN) have been prevalent neural architectures in self-supervised image denoising (SSID). However, most existing BSNs are conducted with convolution layers. Although transformers have shown the potential to overcome the limitations of convolutions in many image restoration tasks, the attention mechanisms may violate the blind-spot requirement, thereby restricting their applicability in BSN. To this end, we propose to analyze and redesign the channel and spatial attentions to meet the blind-spot requirement. Specifically, channel self-attention may leak the blind-spot information in multi-scale architectures, since the downsampling shuffles the spatial feature into channel dimensions. To alleviate this problem, we divide the channel into several groups and perform channel attention separately. For spatial selfattention, we apply an elaborate mask to the attention matrix to restrict and mimic the receptive field of dilated convolution. Based on the redesigned channel and window attentions, we build a Transformer-based Blind-Spot Network (TBSN), which shows strong local fitting and global perspective abilities. Furthermore, we introduce a knowledge distillation strategy that distills TBSN into smaller denoisers to improve computational efficiency while maintaining performance. Extensive experiments on real-world image denoising datasets show that TBSN largely extends the receptive field and exhibits favorable performance against state-of-theart SSID methods.
Neural networks encounter the challenge of Catastrophic Forgetting (CF) in continual learning, where new task learning interferes with previously learned knowledge. Existing data fine-tuning and regularization methods necessitate task identity information during inference and cannot eliminate interference among different tasks, while soft parameter sharing approaches encounter the problem of an increasing model parameter size. To tackle these challenges, we propose the Remembering Transformer, inspired by the brain's Complementary Learning Systems (CLS). Remembering Transformer employs a mixture-of-adapters architecture and a generative model-based novelty detection mechanism in a pretrained Transformer to alleviate CF. Remembering Transformer dynamically routes task data to the most relevant adapter with enhanced parameter efficiency based on knowledge distillation. We conducted extensive experiments, including ablation studies on the novelty detection mechanism and model capacity of the mixture-of-adapters, in a broad range of class-incremental split tasks and permutation tasks. Our approach demonstrated SOTA performance surpassing the second-best method by 15.90% in the split tasks, reducing the memory footprint from 11.18M to 0.22M in the five splits CIFAR10 task.
Despite the widespread applications of machine learning force fields (MLFF) in solids and small molecules, there is a notable gap in applying MLFF to simulate liquid electrolyte, a critical component of the current commercial lithium-ion battery. In this work, we introduce BAMBOO (\textbf{B}yteDance \textbf{A}I \textbf{M}olecular Simulation \textbf{Boo}ster), a predictive framework for molecular dynamics (MD) simulations, with a demonstration of its capability in the context of liquid electrolyte for lithium batteries. We design a physics-inspired graph equivariant transformer architecture as the backbone of BAMBOO to learn from quantum mechanical simulations. Additionally, we introduce an ensemble knowledge distillation approach and apply it to MLFFs to reduce the fluctuation of observations from MD simulations. Finally, we propose a density alignment algorithm to align BAMBOO with experimental measurements. BAMBOO demonstrates state-of-the-art accuracy in predicting key electrolyte properties such as density, viscosity, and ionic conductivity across various solvents and salt combinations. The current model, trained on more than 15 chemical species, achieves the average density error of 0.01 g/cm$^3$ on various compositions compared with experiment.
This paper studies retrieval-augmented approaches for personalizing large language models (LLMs), which potentially have a substantial impact on various applications and domains. We propose the first attempt to optimize the retrieval models that deliver a limited number of personal documents to large language models for the purpose of personalized generation. We develop two optimization algorithms that solicit feedback from the downstream personalized generation tasks for retrieval optimization--one based on reinforcement learning whose reward function is defined using any arbitrary metric for personalized generation and another based on knowledge distillation from the downstream LLM to the retrieval model. This paper also introduces a pre- and post-generation retriever selection model that decides what retriever to choose for each LLM input. Extensive experiments on diverse tasks from the language model personalization (LaMP) benchmark reveal statistically significant improvements in six out of seven datasets.
The domain of computer vision has experienced significant advancements in facial-landmark detection, becoming increasingly essential across various applications such as augmented reality, facial recognition, and emotion analysis. Unlike object detection or semantic segmentation, which focus on identifying objects and outlining boundaries, faciallandmark detection aims to precisely locate and track critical facial features. However, deploying deep learning-based facial-landmark detection models on embedded systems with limited computational resources poses challenges due to the complexity of facial features, especially in dynamic settings. Additionally, ensuring robustness across diverse ethnicities and expressions presents further obstacles. Existing datasets often lack comprehensive representation of facial nuances, particularly within populations like those in Taiwan. This paper introduces a novel approach to address these challenges through the development of a knowledge distillation method. By transferring knowledge from larger models to smaller ones, we aim to create lightweight yet powerful deep learning models tailored specifically for facial-landmark detection tasks. Our goal is to design models capable of accurately locating facial landmarks under varying conditions, including diverse expressions, orientations, and lighting environments. The ultimate objective is to achieve high accuracy and real-time performance suitable for deployment on embedded systems. This method was successfully implemented and achieved a top 6th place finish out of 165 participants in the IEEE ICME 2024 PAIR competition.
Vision-based crack detection faces deployment challenges due to the size of robust models and edge device limitations. These can be addressed with lightweight models trained with knowledge distillation (KD). However, state-of-the-art (SOTA) KD methods compromise anti-noise robustness. This paper develops Robust Feature Knowledge Distillation (RFKD), a framework to improve robustness while retaining the precision of light models for crack segmentation. RFKD distils knowledge from a teacher model's logit layers and intermediate feature maps while leveraging mixed clean and noisy images to transfer robust patterns to the student model, improving its precision, generalisation, and anti-noise performance. To validate the proposed RFKD, a lightweight crack segmentation model, PoolingCrack Tiny (PCT), with only 0.5 M parameters, is also designed and used as the student to run the framework. The results show a significant enhancement in noisy images, with RFKD reaching a 62% enhanced mean Dice score (mDS) compared to SOTA KD methods.
Contrastive Language-Image Pre-training (CLIP) has been shown to improve zero-shot generalization capabilities of language and vision models. In this paper, we extend CLIP for efficient knowledge distillation, by utilizing embeddings as teachers. Typical knowledge distillation frameworks require running forward passes through a teacher model, which is often prohibitive in the case of billion or trillion parameter teachers. In these cases, using only the embeddings of the teacher models to guide the distillation can yield significant computational savings. Our preliminary findings show that CLIP-based knowledge distillation with embeddings can outperform full scale knowledge distillation using $9\times$ less memory and $8\times$ less training time. Code available at: https://github.com/lnairGT/CLIP-Distillation/
Compression techniques have been crucial in advancing machine learning by enabling efficient training and deployment of large-scale language models. However, these techniques have received limited attention in the context of low-resource language models, which are trained on even smaller amounts of data and under computational constraints, a scenario known as the "low-resource double-bind." This paper investigates the effectiveness of pruning, knowledge distillation, and quantization on an exclusively low-resourced, small-data language model, AfriBERTa. Through a battery of experiments, we assess the effects of compression on performance across several metrics beyond accuracy. Our study provides evidence that compression techniques significantly improve the efficiency and effectiveness of small-data language models, confirming that the prevailing beliefs regarding the effects of compression on large, heavily parameterized models hold true for less-parameterized, small-data models.
Radiologists highly desire fully automated versatile AI for medical imaging interpretation. However, the lack of extensively annotated large-scale multi-disease datasets has hindered the achievement of this goal. In this paper, we explore the feasibility of leveraging language as a naturally high-quality supervision for chest CT imaging. In light of the limited availability of image-report pairs, we bootstrap the understanding of 3D chest CT images by distilling chest-related diagnostic knowledge from an extensively pre-trained 2D X-ray expert model. Specifically, we propose a language-guided retrieval method to match each 3D CT image with its semantically closest 2D X-ray image, and perform pair-wise and semantic relation knowledge distillation. Subsequently, we use contrastive learning to align images and reports within the same patient while distinguishing them from the other patients. However, the challenge arises when patients have similar semantic diagnoses, such as healthy patients, potentially confusing if treated as negatives. We introduce a robust contrastive learning that identifies and corrects these false negatives. We train our model with over 12,000 pairs of chest CT images and radiology reports. Extensive experiments across multiple scenarios, including zero-shot learning, report generation, and fine-tuning processes, demonstrate the model's feasibility in interpreting chest CT images.
Monocular 3D object detection (Mono3D) holds noteworthy promise for autonomous driving applications owing to the cost-effectiveness and rich visual context of monocular camera sensors. However, depth ambiguity poses a significant challenge, as it requires extracting precise 3D scene geometry from a single image, resulting in suboptimal performance when transferring knowledge from a LiDAR-based teacher model to a camera-based student model. To facilitate effective distillation, we introduce Monocular Teaching Assistant Knowledge Distillation (MonoTAKD), which proposes a camera-based teaching assistant (TA) model to transfer robust 3D visual knowledge to the student model, leveraging the smaller feature representation gap. Additionally, we define 3D spatial cues as residual features that capture the differences between the teacher and the TA models. We then leverage these cues to improve the student model's 3D perception capabilities. Experimental results show that our MonoTAKD achieves state-of-the-art performance on the KITTI3D dataset. Furthermore, we evaluate the performance on nuScenes and KITTI raw datasets to demonstrate the generalization of our model to multi-view 3D and unsupervised data settings. Our code is available at https://github.com/hoiliu-0801/MonoTAKD.
Score distillation sampling~(SDS) has been widely adopted to overcome the absence of unseen views in reconstructing 3D objects from a \textbf{single} image. It leverages pre-trained 2D diffusion models as teacher to guide the reconstruction of student 3D models. Despite their remarkable success, SDS-based methods often encounter geometric artifacts and texture saturation. We find out the crux is the overlooked indiscriminate treatment of diffusion time-steps during optimization: it unreasonably treats the student-teacher knowledge distillation to be equal at all time-steps and thus entangles coarse-grained and fine-grained modeling. Therefore, we propose the Diffusion Time-step Curriculum one-image-to-3D pipeline (DTC123), which involves both the teacher and student models collaborating with the time-step curriculum in a coarse-to-fine manner. Extensive experiments on NeRF4, RealFusion15, GSO and Level50 benchmark demonstrate that DTC123 can produce multi-view consistent, high-quality, and diverse 3D assets. Codes and more generation demos will be released in https://github.com/yxymessi/DTC123.
With the power of large language models (LLMs), open-ended embodied agents can flexibly understand human instructions, generate interpretable guidance strategies, and output executable actions. Nowadays, Multi-modal Language Models~(MLMs) integrate multi-modal signals into LLMs, further bringing richer perception to entity agents and allowing embodied agents to perceive world-understanding tasks more delicately. However, existing works: 1) operate independently by agents, each containing multiple LLMs, from perception to action, resulting in gaps between complex tasks and execution; 2) train MLMs on static data, struggling with dynamics in open-ended scenarios; 3) input prior knowledge directly as prompts, suppressing application flexibility. We propose STEVE-2, a hierarchical knowledge distillation framework for open-ended embodied tasks, characterized by 1) a hierarchical system for multi-granular task division, 2) a mirrored distillation method for parallel simulation data, and 3) an extra expert model for bringing additional knowledge into parallel simulation. After distillation, embodied agents can complete complex, open-ended tasks without additional expert guidance, utilizing the performance and knowledge of a versatile MLM. Extensive evaluations on navigation and creation tasks highlight the superior performance of STEVE-2 in open-ended tasks, with $1.4 \times$ - $7.3 \times$ in performance.
Knowledge distillation (KD) has become a widely used technique in the field of model compression, which aims to transfer knowledge from a large teacher model to a lightweight student model for efficient network development. In addition to the supervision of ground truth, the vanilla KD method regards the predictions of the teacher as soft labels to supervise the training of the student model. Based on vanilla KD, various approaches have been developed to further improve the performance of the student model. However, few of these previous methods have considered the reliability of the supervision from teacher models. Supervision from erroneous predictions may mislead the training of the student model. This paper therefore proposes to tackle this problem from two aspects: Label Revision to rectify the incorrect supervision and Data Selection to select appropriate samples for distillation to reduce the impact of erroneous supervision. In the former, we propose to rectify the teacher's inaccurate predictions using the ground truth. In the latter, we introduce a data selection technique to choose suitable training samples to be supervised by the teacher, thereby reducing the impact of incorrect predictions to some extent. Experiment results demonstrate the effectiveness of our proposed method, and show that our method can be combined with other distillation approaches, improving their performance.
In NLP, Event Coreference Resolution (ECR) is the task of connecting event clusters that refer to the same underlying real-life event, usually via neural systems. In this work, we investigate using abductive free-text rationales (FTRs) generated by modern autoregressive LLMs as distant supervision of smaller student models for cross-document coreference (CDCR) of events. We implement novel rationale-oriented event clustering and knowledge distillation methods for event coreference scoring that leverage enriched information from the FTRs for improved CDCR without additional annotation or expensive document clustering. Our model using coreference specific knowledge distillation achieves SOTA B3 F1 on the ECB+ and GVC corpora and we establish a new baseline on the AIDA Phase 1 corpus. Our code can be found at https://github.com/csu-signal/llama_cdcr