The growth of digital documents presents significant challenges in efficient management and knowledge extraction. Traditional methods often struggle with complex documents, leading to issues such as hallucinations and high latency in responses from Large Language Models (LLMs). ZeroG, an innovative approach, significantly mitigates these challenges by leveraging knowledge distillation and prompt tuning to enhance model performance. ZeroG utilizes a smaller model that replicates the behavior of a larger teacher model, ensuring contextually relevant and grounded responses, by employing a black-box distillation approach, it creates a distilled dataset without relying on intermediate features, optimizing computational efficiency. This method significantly enhances accuracy and reduces response times, providing a balanced solution for modern document management. Incorporating advanced techniques for document ingestion and metadata utilization, ZeroG improves the accuracy of question-and-answer systems. The integration of graph databases and robust metadata management further streamlines information retrieval, allowing for precise and context-aware responses. By transforming how organizations interact with complex data, ZeroG enhances productivity and user experience, offering a scalable solution for the growing demands of digital document management.
In this paper, we propose a novel LLM-Neo framework that efficiently transfers knowledge from a large language model (LLM) teacher to a compact student. Initially, we revisit the knowledge distillation (KD) and low-rank adaption (LoRA), and argue that they share the same paradigm. Inspired by this observation, we explore the strategy that combines LoRA and KD to enhance the efficiency of knowledge transfer. We first summarize some guidelines for this design and further develop the LLM-Neo. Experimental results on compressing Llama 2 and Llama 3 show that LLM-Neo outperforms various baselines. Further analysis demonstrates the robustness of the proposed LLM-Neo on variants of LoRA. The trained models have been available at \href{https://huggingface.co/collections/yang31210999/llm-neo-66e3c882f5579b829ff57eba}{this repository}.
In this paper, we question if well pre-trained vision transformer (ViT) models could be used as teachers that exhibit scalable properties to advance cross architecture knowledge distillation (KD) research, in the context of using large-scale datasets for evaluation. To make this possible, our analysis underlines the importance of seeking effective strategies to align (1) feature computing paradigm differences, (2) model scale differences, and (3) knowledge density differences. By combining three coupled components namely cross attention projector, dual-view feature mimicking and teacher parameter perception tailored to address the above problems, we present a simple and effective KD method, called ScaleKD. Our method can train student backbones that span across a variety of convolutional neural network (CNN), multi-layer perceptron (MLP), and ViT architectures on image classification datasets, achieving state-of-the-art distillation performance. For instance, taking a well pre-trained Swin-L as the teacher model, our method gets 75.15%|82.03%|84.16%|78.63%|81.96%|83.93%|83.80%|85.53% top-1 accuracies for MobileNet-V1|ResNet-50|ConvNeXt-T|Mixer-S/16|Mixer-B/16|ViT-S/16|Swin-T|ViT-B/16 models trained on ImageNet-1K dataset from scratch, showing 3.05%|3.39%|2.02%|4.61%|5.52%|4.03%|2.62%|3.73% absolute gains to the individually trained counterparts. Intriguingly, when scaling up the size of teacher models or their pre-training datasets, our method showcases the desired scalable properties, bringing increasingly larger gains to student models. The student backbones trained by our method transfer well on downstream MS-COCO and ADE20K datasets. More importantly, our method could be used as a more efficient alternative to the time-intensive pre-training paradigm for any target student model if a strong pre-trained ViT is available, reducing the amount of viewed training samples up to 195x.
Incremental graph learning has gained significant attention for its ability to address the catastrophic forgetting problem in graph representation learning. However, traditional methods often rely on a large number of labels for node classification, which is impractical in real-world applications. This makes few-shot incremental learning on graphs a pressing need. Current methods typically require extensive training samples from meta-learning to build memory and perform intensive fine-tuning of GNN parameters, leading to high memory consumption and potential loss of previously learned knowledge. To tackle these challenges, we introduce Mecoin, an efficient method for building and maintaining memory. Mecoin employs Structured Memory Units to cache prototypes of learned categories, as well as Memory Construction Modules to update these prototypes for new categories through interactions between the nodes and the cached prototypes. Additionally, we have designed a Memory Representation Adaptation Module to store probabilities associated with each class prototype, reducing the need for parameter fine-tuning and lowering the forgetting rate. When a sample matches its corresponding class prototype, the relevant probabilities are retrieved from the MRaM. Knowledge is then distilled back into the GNN through a Graph Knowledge Distillation Module, preserving the model's memory. We analyze the effectiveness of Mecoin in terms of generalization error and explore the impact of different distillation strategies on model performance through experiments and VC-dimension analysis. Compared to other related works, Mecoin shows superior performance in accuracy and forgetting rate. Our code is publicly available on the https://github.com/Arvin0313/Mecoin-GFSCIL.git .
Large Language Models (LLMs) face significant challenges at inference time due to their high computational demands. To address this, we present Performance-Guided Knowledge Distillation (PGKD), a cost-effective and high-throughput solution for production text classification applications. PGKD utilizes teacher-student Knowledge Distillation to distill the knowledge of LLMs into smaller, task-specific models. PGKD establishes an active learning routine between the student model and the LLM; the LLM continuously generates new training data leveraging hard-negative mining, student model validation performance, and early-stopping protocols to inform the data generation. By employing a cyclical, performance-aware approach tailored for highly multi-class, sparsely annotated datasets prevalent in industrial text classification, PGKD effectively addresses training challenges and outperforms traditional BERT-base models and other knowledge distillation methods on several multi-class classification datasets. Additionally, cost and latency benchmarking reveals that models fine-tuned with PGKD are up to 130X faster and 25X less expensive than LLMs for inference on the same classification task. While PGKD is showcased for text classification tasks, its versatile framework can be extended to any LLM distillation task, including language generation, making it a powerful tool for optimizing performance across a wide range of AI applications.
This paper describes Asterisk, a compact GPT-based model for generating text embeddings. The model uses a minimalist architecture with two layers, two attention heads, and 256 embedding dimensions. By applying knowledge distillation from larger pretrained models, we explore the trade-offs between model size and performance while minimizing computational and memory requirements. The model is primarily evaluated and optimized for classification tasks, with experimental results showing its moderate performance in zero-shot classification across various downstream applications. With additional configuration, the model performance can approach or even surpass that of larger architectures on specific classification tasks.
As we move towards a mixed-traffic scenario of Autonomous vehicles (AVs) and Human-driven vehicles (HDVs), understanding the car-following behaviour is important to improve traffic efficiency and road safety. Using a real-world trajectory dataset, this study uses descriptive and statistical analysis to investigate the car-following behaviours of three vehicle pairs: HDV-AV, AV-HDV and HDV-HDV in mixed traffic. The ANOVA test showed that car-following behaviours across different vehicle pairs are statistically significant (p-value < 0.05). We also introduce a data-driven Knowledge Distillation Neural Network (KDNN) model for predicting car-following behaviour in terms of speed. The KDNN model demonstrates comparable predictive accuracy to its teacher network, a Long Short-Term Memory (LSTM) network, and outperforms both the standalone student network, a Multilayer Perceptron (MLP), and traditional physics-based models like the Gipps model. Notably, the KDNN model better prevents collisions, measured by minimum Time-to-Collision (TTC), and operates with lower computational power, making it ideal for AVs or driving simulators requiring efficient computing.
Lifelong few-shot customization for text-to-image diffusion aims to continually generalize existing models for new tasks with minimal data while preserving old knowledge. Current customization diffusion models excel in few-shot tasks but struggle with catastrophic forgetting problems in lifelong generations. In this study, we identify and categorize the catastrophic forgetting problems into two folds: relevant concepts forgetting and previous concepts forgetting. To address these challenges, we first devise a data-free knowledge distillation strategy to tackle relevant concepts forgetting. Unlike existing methods that rely on additional real data or offline replay of original concept data, our approach enables on-the-fly knowledge distillation to retain the previous concepts while learning new ones, without accessing any previous data. Second, we develop an In-Context Generation (ICGen) paradigm that allows the diffusion model to be conditioned upon the input vision context, which facilitates the few-shot generation and mitigates the issue of previous concepts forgetting. Extensive experiments show that the proposed Lifelong Few-Shot Diffusion (LFS-Diffusion) method can produce high-quality and accurate images while maintaining previously learned knowledge.
Learned sparse retrieval, which can efficiently perform retrieval through mature inverted-index engines, has garnered growing attention in recent years. Particularly, the inference-free sparse retrievers are attractive as they eliminate online model inference in the retrieval phase thereby avoids huge computational cost, offering reasonable throughput and latency. However, even the state-of-the-art (SOTA) inference-free sparse models lag far behind in terms of search relevance when compared to both sparse and dense siamese models. Towards competitive search relevance for inference-free sparse retrievers, we argue that they deserve dedicated training methods other than using same ones with siamese encoders. In this paper, we propose two different approaches for performance improvement. First, we introduce the IDF-aware FLOPS loss, which introduces Inverted Document Frequency (IDF) to the sparsification of representations. We find that it mitigates the negative impact of the FLOPS regularization on search relevance, allowing the model to achieve a better balance between accuracy and efficiency. Moreover, we propose a heterogeneous ensemble knowledge distillation framework that combines siamese dense and sparse retrievers to generate supervisory signals during the pre-training phase. The ensemble framework of dense and sparse retriever capitalizes on their strengths respectively, providing a strong upper bound for knowledge distillation. To concur the diverse feedback from heterogeneous supervisors, we normalize and then aggregate the outputs of the teacher models to eliminate score scale differences. On the BEIR benchmark, our model outperforms existing SOTA inference-free sparse model by \textbf{3.3 NDCG@10 score}. It exhibits search relevance comparable to siamese sparse retrievers and client-side latency only \textbf{1.1x that of BM25}.
We present GazeGen, a user interaction system that generates visual content (images and videos) for locations indicated by the user's eye gaze. GazeGen allows intuitive manipulation of visual content by targeting regions of interest with gaze. Using advanced techniques in object detection and generative AI, GazeGen performs gaze-controlled image adding/deleting, repositioning, and surface style changes of image objects, and converts static images into videos. Central to GazeGen is the DFT Gaze (Distilled and Fine-Tuned Gaze) agent, an ultra-lightweight model with only 281K parameters, performing accurate real-time gaze predictions tailored to individual users' eyes on small edge devices. GazeGen is the first system to combine visual content generation with real-time gaze estimation, made possible exclusively by DFT Gaze. This real-time gaze estimation enables various visual content generation tasks, all controlled by the user's gaze. The input for DFT Gaze is the user's eye images, while the inputs for visual content generation are the user's view and the predicted gaze point from DFT Gaze. To achieve efficient gaze predictions, we derive the small model from a large model (10x larger) via novel knowledge distillation and personal adaptation techniques. We integrate knowledge distillation with a masked autoencoder, developing a compact yet powerful gaze estimation model. This model is further fine-tuned with Adapters, enabling highly accurate and personalized gaze predictions with minimal user input. DFT Gaze ensures low-latency and precise gaze tracking, supporting a wide range of gaze-driven tasks. We validate the performance of DFT Gaze on AEA and OpenEDS2020 benchmarks, demonstrating low angular gaze error and low latency on the edge device (Raspberry Pi 4). Furthermore, we describe applications of GazeGen, illustrating its versatility and effectiveness in various usage scenarios.
Fluoroscopy is critical for real-time X-ray visualization in medical imaging. However, low-dose images are compromised by noise, potentially affecting diagnostic accuracy. Noise reduction is crucial for maintaining image quality, especially given such challenges as motion artifacts and the limited availability of clean data in medical imaging. To address these issues, we propose an unsupervised training framework for dynamic context-aware denoising of fluoroscopy image sequences. First, we train the multi-scale recurrent attention U-Net (MSR2AU-Net) without requiring clean data to address the initial noise. Second, we incorporate a knowledge distillation-based uncorrelated noise suppression module and a recursive filtering-based correlated noise suppression module enhanced with motion compensation to further improve motion compensation and achieve superior denoising performance. Finally, we introduce a novel approach by combining these modules with a pixel-wise dynamic object motion cross-fusion matrix, designed to adapt to motion, and an edge-preserving loss for precise detail retention. To validate the proposed method, we conducted extensive numerical experiments on medical image datasets, including 3500 fluoroscopy images from dynamic phantoms (2,400 images for training, 1,100 for testing) and 350 clinical images from a spinal surgery patient. Moreover, we demonstrated the robustness of our approach across different imaging modalities by testing it on the publicly available 2016 Low Dose CT Grand Challenge dataset, using 4,800 images for training and 1,136 for testing. The results demonstrate that the proposed approach outperforms state-of-the-art unsupervised algorithms in both visual quality and quantitative evaluation while achieving comparable performance to well-established supervised learning methods across low-dose fluoroscopy and CT imaging.
Logic synthesis is a crucial phase in the circuit design process, responsible for transforming hardware description language (HDL) designs into optimized netlists. However, traditional logic synthesis methods are computationally intensive, restricting their iterative use in refining chip designs. Recent advancements in large language models (LLMs), particularly those fine-tuned on programming languages, present a promising alternative. This work proposes augmenting LLMs with predictor networks trained to estimate circuit quality directly from HDL code. To enhance performance, the model is regularized using embeddings from graph neural networks (GNNs) trained on Look-Up Table (LUT) graphs, thereby incorporating lower-level circuit insights. The proposed method demonstrates superior performance compared to existing graph-based RTL-level estimation techniques on the established benchmark OpenABCD, while providing instant feedback on HDL code quality.
In safeguarding mission-critical systems, such as Unmanned Aerial Vehicles (UAVs), preserving the privacy of path trajectories during navigation is paramount. While the combination of Reinforcement Learning (RL) and Fully Homomorphic Encryption (FHE) holds promise, the computational overhead of FHE presents a significant challenge. This paper proposes an innovative approach that leverages Knowledge Distillation to enhance the practicality of secure UAV navigation. By integrating RL and FHE, our framework addresses vulnerabilities to adversarial attacks while enabling real-time processing of encrypted UAV camera feeds, ensuring data security. To mitigate FHE's latency, Knowledge Distillation is employed to compress the network, resulting in an impressive 18x speedup without compromising performance, as evidenced by an R-squared score of 0.9499 compared to the original model's score of 0.9631. Our methodology underscores the feasibility of processing encrypted data for UAV navigation tasks, emphasizing security alongside performance efficiency and timely processing. These findings pave the way for deploying autonomous UAVs in sensitive environments, bolstering their resilience against potential security threats.
Federated learning is a distributed machine learning paradigm designed to protect data privacy. However, data heterogeneity across various clients results in catastrophic forgetting, where the model rapidly forgets previous knowledge while acquiring new knowledge. To address this challenge, personalized federated learning has emerged to customize a personalized model for each client. However, the inherent limitation of this mechanism is its excessive focus on personalization, potentially hindering the generalization of those models. In this paper, we present a novel personalized federated learning method that uses global and historical models as teachers and the local model as the student to facilitate comprehensive knowledge distillation. The historical model represents the local model from the last round of client training, containing historical personalized knowledge, while the global model represents the aggregated model from the last round of server aggregation, containing global generalized knowledge. By applying knowledge distillation, we effectively transfer global generalized knowledge and historical personalized knowledge to the local model, thus mitigating catastrophic forgetting and enhancing the general performance of personalized models. Extensive experimental results demonstrate the significant advantages of our method.
This study presents a transformer-based approach for fault-tolerant control in fixed-wing Unmanned Aerial Vehicles (UAVs), designed to adapt in real time to dynamic changes caused by structural damage or actuator failures. Unlike traditional Flight Control Systems (FCSs) that rely on classical control theory and struggle under severe alterations in dynamics, our method directly maps outer-loop reference values -- altitude, heading, and airspeed -- into control commands using the in-context learning and attention mechanisms of transformers, thus bypassing inner-loop controllers and fault-detection layers. Employing a teacher-student knowledge distillation framework, the proposed approach trains a student agent with partial observations by transferring knowledge from a privileged expert agent with full observability, enabling robust performance across diverse failure scenarios. Experimental results demonstrate that our transformer-based controller outperforms industry-standard FCS and state-of-the-art reinforcement learning (RL) methods, maintaining high tracking accuracy and stability in nominal conditions and extreme failure cases, highlighting its potential for enhancing UAV operational safety and reliability.
Developing accurate and efficient detectors for drone imagery is challenging due to the inherent complexity of aerial scenes. While some existing methods aim to achieve high accuracy by utilizing larger models, their computational cost is prohibitive for drones. Recently, Knowledge Distillation (KD) has shown promising potential for maintaining satisfactory accuracy while significantly compressing models in general object detection. Considering the advantages of KD, this paper presents the first attempt to adapt it to object detection on drone imagery and addresses two intrinsic issues: (1) low foreground-background ratio and (2) small instances and complex backgrounds, which lead to inadequate training, resulting insufficient distillation. Therefore, we propose a task-wise Lightweight Mutual Lifting (Light-ML) module with a Centerness-based Instance-aware Distillation (CID) strategy. The Light-ML module mutually harmonizes the classification and localization branches by channel shuffling and convolution, integrating teacher supervision across different tasks during back-propagation, thus facilitating training the student model. The CID strategy extracts valuable regions surrounding instances through the centerness of proposals, enhancing distillation efficacy. Experiments on the VisDrone, UAVDT, and COCO benchmarks demonstrate that the proposed approach promotes the accuracies of existing state-of-the-art KD methods with comparable computational requirements. Codes will be available upon acceptance.
Existing Multimodal Large Language Models (MLLMs) and Visual Language Pretrained Models (VLPMs) have shown remarkable performances in the general Visual Question Answering (VQA). However, these models struggle with VQA questions that require external commonsense knowledge due to the challenges in generating high-quality prompts and the high computational costs of fine-tuning. In this work, we propose a novel graph-based multimodal commonsense knowledge distillation framework that constructs a unified relational graph over commonsense knowledge, visual objects and questions through a Graph Convolutional Network (GCN) following a teacher-student environment. This proposed framework is flexible with any type of teacher and student models without further fine-tuning, and has achieved competitive performances on the ScienceQA dataset.
Neural approaches to ranking based on pre-trained language models are highly effective in ad-hoc search. However, the computational expense of these models can limit their application. As such, a process known as knowledge distillation is frequently applied to allow a smaller, efficient model to learn from an effective but expensive model. A key example of this is the distillation of expensive API-based commercial Large Language Models into smaller production-ready models. However, due to the opacity of training data and processes of most commercial models, one cannot ensure that a chosen test collection has not been observed previously, creating the potential for inadvertent data contamination. We, therefore, investigate the effect of a contaminated teacher model in a distillation setting. We evaluate several distillation techniques to assess the degree to which contamination occurs during distillation. By simulating a ``worst-case'' setting where the degree of contamination is known, we find that contamination occurs even when the test data represents a small fraction of the teacher's training samples. We, therefore, encourage caution when training using black-box teacher models where data provenance is ambiguous.
A myriad of factors affect large scale ads delivery systems and influence both user experience and revenue. One such factor is proactive detection and calibration of seasonal advertisements to help with increasing conversion and user satisfaction. In this paper, we present Proactive Detection and Calibration of Seasonal Advertisements (PDCaSA), a research problem that is of interest for the ads ranking and recommendation community, both in the industrial setting as well as in research. Our paper provides detailed guidelines from various angles of this problem tested in, and motivated by a large-scale industrial ads ranking system. We share our findings including the clear statement of the problem and its motivation rooted in real-world systems, evaluation metrics, and sheds lights to the existing challenges, lessons learned, and best practices of data annotation and machine learning modeling to tackle this problem. Lastly, we present a conclusive solution we took during this research exploration: to detect seasonality, we leveraged Multimodal LLMs (MLMs) which on our in-house benchmark achieved 0.97 top F1 score. Based on our findings, we envision MLMs as a teacher for knowledge distillation, a machine labeler, and a part of the ensembled and tiered seasonality detection system, which can empower ads ranking systems with enriched seasonal information.
This study presents an innovative dynamic weighting knowledge distillation (KD) framework tailored for efficient Earth observation (EO) image classification (IC) in resource-constrained settings. Utilizing EfficientViT and MobileViT as teacher models, this framework enables lightweight student models, particularly ResNet8 and ResNet16, to surpass 90% in accuracy, precision, and recall, adhering to the stringent confidence thresholds necessary for reliable classification tasks. Unlike conventional KD methods that rely on static weight distribution, our adaptive weighting mechanism responds to each teacher model's confidence, allowing student models to prioritize more credible sources of knowledge dynamically. Remarkably, ResNet8 delivers substantial efficiency gains, achieving a 97.5% reduction in parameters, a 96.7% decrease in FLOPs, an 86.2% cut in power consumption, and a 63.5% increase in inference speed over MobileViT. This significant optimization of complexity and resource demands establishes ResNet8 as an optimal candidate for EO tasks, combining robust performance with feasibility in deployment. The confidence-based, adaptable KD approach underscores the potential of dynamic distillation strategies to yield high-performing, resource-efficient models tailored for satellite-based EO applications. The reproducible code is accessible on our GitHub repository.