Federated Distillation (FD) offers an innovative approach to distributed machine learning, leveraging knowledge distillation for efficient and flexible cross-device knowledge transfer without necessitating the upload of extensive model parameters to a central server. While FD has gained popularity, its vulnerability to poisoning attacks remains underexplored. To address this gap, we previously introduced FDLA (Federated Distillation Logits Attack), a method that manipulates logits communication to mislead and degrade the performance of client models. However, the impact of FDLA on participants with different identities and the effects of malicious modifications at various stages of knowledge transfer remain unexplored. To this end, we present PCFDLA (Peak-Controlled Federated Distillation Logits Attack), an advanced and more stealthy logits poisoning attack method for FD. PCFDLA enhances the effectiveness of FDLA by carefully controlling the peak values of logits to create highly misleading yet inconspicuous modifications. Furthermore, we introduce a novel metric for better evaluating attack efficacy, demonstrating that PCFDLA maintains stealth while being significantly more disruptive to victim models compared to its predecessors. Experimental results across various datasets confirm the superior impact of PCFDLA on model accuracy, solidifying its potential threat in federated distillation systems.
Novel Class Discovery (NCD) involves identifying new categories within unlabeled data by utilizing knowledge acquired from previously established categories. However, existing NCD methods often struggle to maintain a balance between the performance of old and new categories. Discovering unlabeled new categories in a class-incremental way is more practical but also more challenging, as it is frequently hindered by either catastrophic forgetting of old categories or an inability to learn new ones. Furthermore, the implementation of NCD on continuously scalable graph-structured data remains an under-explored area. In response to these challenges, we introduce for the first time a more practical NCD scenario for node classification (i.e., NC-NCD), and propose a novel self-training framework with prototype replay and distillation called SWORD, adopted to our NC-NCD setting. Our approach enables the model to cluster unlabeled new category nodes after learning labeled nodes while preserving performance on old categories without reliance on old category nodes. SWORD achieves this by employing a self-training strategy to learn new categories and preventing the forgetting of old categories through the joint use of feature prototypes and knowledge distillation. Extensive experiments on four common benchmarks demonstrate the superiority of SWORD over other state-of-the-art methods.
Despite the advanced intelligence abilities of large language models (LLMs) in various applications, they still face significant computational and storage demands. Knowledge Distillation (KD) has emerged as an effective strategy to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a high-performing LLM (i.e., the teacher model). Prevailing techniques in LLM distillation typically use a black-box model API to generate high-quality pretrained and aligned datasets, or utilize white-box distillation by altering the loss function to better transfer knowledge from the teacher LLM. However, these methods ignore the knowledge differences between the student and teacher LLMs across domains. This results in excessive focus on domains with minimal performance gaps and insufficient attention to domains with large gaps, reducing overall performance. In this paper, we introduce a new LLM distillation framework called DDK, which dynamically adjusts the composition of the distillation dataset in a smooth manner according to the domain performance differences between the teacher and student models, making the distillation process more stable and effective. Extensive evaluations show that DDK significantly improves the performance of student models, outperforming both continuously pretrained baselines and existing knowledge distillation methods by a large margin.
Deep learning models have achieved tremendous success in most of the industries in recent years. The evolution of these models has also led to an increase in the model size and energy requirement, making it difficult to deploy in production on low compute devices. An increase in the number of connected devices around the world warrants compressed models that can be easily deployed at the local devices with low compute capacity and power accessibility. A wide range of solutions have been proposed by different researchers to reduce the size and complexity of such models, prominent among them are, Weight Quantization, Parameter Pruning, Network Pruning, low-rank representation, weights sharing, neural architecture search, knowledge distillation etc. In this research work, we investigate the performance impacts on various trained deep learning models, compressed using quantization and pruning techniques. We implemented both, quantization and pruning, compression techniques on popular deep learning models used in the image classification, object detection, language models and generative models-based problem statements. We also explored performance of various large language models (LLMs) after quantization and low rank adaptation. We used the standard evaluation metrics (model's size, accuracy, and inference time) for all the related problem statements and concluded this paper by discussing the challenges and future work.
Knowledge distillation (KD) is a model compression method that entails training a compact student model to emulate the performance of a more complex teacher model. However, the architectural capacity gap between the two models limits the effectiveness of knowledge transfer. Addressing this issue, previous works focused on customizing teacher-student pairs to improve compatibility, a computationally expensive process that needs to be repeated every time either model changes. Hence, these methods are impractical when a teacher model has to be compressed into different student models for deployment on multiple hardware devices with distinct resource constraints. In this work, we propose Generic Teacher Network (GTN), a one-off KD-aware training to create a generic teacher capable of effectively transferring knowledge to any student model sampled from a given finite pool of architectures. To this end, we represent the student pool as a weight-sharing supernet and condition our generic teacher to align with the capacities of various student architectures sampled from this supernet. Experimental evaluation shows that our method both improves overall KD effectiveness and amortizes the minimal additional training cost of the generic teacher across students in the pool.
Recent studies have demonstrated the significant potential of Large Language Models (LLMs) in generating Register Transfer Level (RTL) code, with notable advancements showcased by commercial models such as GPT-4 and Claude3-Opus. However, these proprietary LLMs often raise concerns regarding privacy and security. While open-source LLMs offer solutions to these concerns, they typically underperform commercial models in RTL code generation tasks, primarily due to the scarcity of high-quality open-source RTL datasets. To address this challenge, we introduce OriGen , a fully open-source framework that incorporates self-reflection capabilities and a novel dataset augmentation methodology for generating high-quality, large-scale RTL code. Our approach employs a code-tocode augmentation technique to enhance the quality of open-source RTL code datasets. Furthermore, OriGen can rectify syntactic errors through a self-reflection process that leverages compiler feedback. Experimental results demonstrate that OriGen significantly outperforms other open-source alternatives in RTL code generation. It surpasses the previous best-performing open-source LLM by 12.8% and even exceeds GPT-4 Turbo in the pass@1 metric on the VerilogEval-Human benchmark. Moreover, OriGen exhibits superior capabilities in self-reflection and error correction, outperforming GPT-4 by 19.9% on a benchmark designed to evaluate self-reflection capabilities.
Weakly supervised video object segmentation (WSVOS) enables the identification of segmentation maps without requiring an extensive training dataset of object masks, relying instead on coarse video labels indicating object presence. Current state-of-the-art methods either require multiple independent stages of processing that employ motion cues or, in the case of end-to-end trainable networks, lack in segmentation accuracy, in part due to the difficulty of learning segmentation maps from videos with transient object presence. This limits the application of WSVOS for semantic annotation of surgical videos where multiple surgical tools frequently move in and out of the field of view, a problem that is more difficult than typically encountered in WSVOS. This paper introduces Video Spatio-Temporal Disentanglement Networks (VDST-Net), a framework to disentangle spatiotemporal information using semi-decoupled knowledge distillation to predict high-quality class activation maps (CAMs). A teacher network designed to resolve temporal conflicts when specifics about object location and timing in the video are not provided works with a student network that integrates information over time by leveraging temporal dependencies. We demonstrate the efficacy of our framework on a public reference dataset and on a more challenging surgical video dataset where objects are, on average, present in less than 60\% of annotated frames. Our method outperforms state-of-the-art techniques and generates superior segmentation masks under video-level weak supervision.
To bridge the gaps between powerful Graph Neural Networks (GNNs) and lightweight Multi-Layer Perceptron (MLPs), GNN-to-MLP Knowledge Distillation (KD) proposes to distill knowledge from a well-trained teacher GNN into a student MLP. In this paper, we revisit the knowledge samples (nodes) in teacher GNNs from the perspective of hardness, and identify that hard sample distillation may be a major performance bottleneck of existing graph KD algorithms. The GNN-to-MLP KD involves two different types of hardness, one student-free knowledge hardness describing the inherent complexity of GNN knowledge, and the other student-dependent distillation hardness describing the difficulty of teacher-to-student distillation. However, most of the existing work focuses on only one of these aspects or regards them as one thing. This paper proposes a simple yet effective Hardness-aware GNN-to-MLP Distillation (HGMD) framework, which decouples the two hardnesses and estimates them using a non-parametric approach. Finally, two hardness-aware distillation schemes (i.e., HGMD-weight and HGMD-mixup) are further proposed to distill hardness-aware knowledge from teacher GNNs into the corresponding nodes of student MLPs. As non-parametric distillation, HGMD does not involve any additional learnable parameters beyond the student MLPs, but it still outperforms most of the state-of-the-art competitors. HGMD-mixup improves over the vanilla MLPs by 12.95% and outperforms its teacher GNNs by 2.48% averaged over seven real-world datasets.
Large language models (LLMs) targeting different deployment scales and sizes are currently produced by training each variant from scratch; this is extremely compute-intensive. In this paper, we investigate if pruning an existing LLM and then re-training it with a fraction (<3%) of the original training data can be a suitable alternative to repeated, full retraining. To this end, we develop a set of practical and effective compression best practices for LLMs that combine depth, width, attention and MLP pruning with knowledge distillation-based retraining; we arrive at these best practices through a detailed empirical exploration of pruning strategies for each axis, methods to combine axes, distillation strategies, and search techniques for arriving at optimal compressed architectures. We use this guide to compress the Nemotron-4 family of LLMs by a factor of 2-4x, and compare their performance to similarly-sized models on a variety of language modeling tasks. Deriving 8B and 4B models from an already pretrained 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch; this results in compute cost savings of 1.8x for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. We have open-sourced Minitron model weights on Huggingface, with corresponding supplementary material including example code available on GitHub.
Data-Free Knowledge Distillation (DFKD) has shown great potential in creating a compact student model while alleviating the dependency on real training data by synthesizing surrogate data. However, prior arts are seldom discussed under distribution shifts, which may be vulnerable in real-world applications. Recent Vision-Language Foundation Models, e.g., CLIP, have demonstrated remarkable performance in zero-shot out-of-distribution generalization, yet consuming heavy computation resources. In this paper, we discuss the extension of DFKD to Vision-Language Foundation Models without access to the billion-level image-text datasets. The objective is to customize a student model for distribution-agnostic downstream tasks with given category concepts, inheriting the out-of-distribution generalization capability from the pre-trained foundation models. In order to avoid generalization degradation, the primary challenge of this task lies in synthesizing diverse surrogate images driven by text prompts. Since not only category concepts but also style information are encoded in text prompts, we propose three novel Prompt Diversification methods to encourage image synthesis with diverse styles, namely Mix-Prompt, Random-Prompt, and Contrastive-Prompt. Experiments on out-of-distribution generalization datasets demonstrate the effectiveness of the proposed methods, with Contrastive-Prompt performing the best.
Generative artificial intelligence has transformed the generation of synthetic data, providing innovative solutions to challenges like data scarcity and privacy, which are particularly critical in fields such as medicine. However, the effective use of this synthetic data to train high-performance models remains a significant challenge. This paper addresses this issue by introducing Knowledge Recycling (KR), a pipeline designed to optimise the generation and use of synthetic data for training downstream classifiers. At the heart of this pipeline is Generative Knowledge Distillation (GKD), the proposed technique that significantly improves the quality and usefulness of the information provided to classifiers through a synthetic dataset regeneration and soft labelling mechanism. The KR pipeline has been tested on a variety of datasets, with a focus on six highly heterogeneous medical image datasets, ranging from retinal images to organ scans. The results show a significant reduction in the performance gap between models trained on real and synthetic data, with models based on synthetic data outperforming those trained on real data in some cases. Furthermore, the resulting models show almost complete immunity to Membership Inference Attacks, manifesting privacy properties missing in models trained with conventional techniques.
Most existing membership inference attacks (MIAs) utilize metrics (e.g., loss) calculated on the model's final state, while recent advanced attacks leverage metrics computed at various stages, including both intermediate and final stages, throughout the model training. Nevertheless, these attacks often process multiple intermediate states of the metric independently, ignoring their time-dependent patterns. Consequently, they struggle to effectively distinguish between members and non-members who exhibit similar metric values, particularly resulting in a high false-positive rate. In this study, we delve deeper into the new membership signals in the black-box scenario. We identify a new, more integrated membership signal: the Pattern of Metric Sequence, derived from the various stages of model training. We contend that current signals provide only partial perspectives of this new signal: the new one encompasses both the model's multiple intermediate and final states, with a greater emphasis on temporal patterns among them. Building upon this signal, we introduce a novel attack method called Sequential-metric based Membership Inference Attack (SeqMIA). Specifically, we utilize knowledge distillation to obtain a set of distilled models representing various stages of the target model's training. We then assess multiple metrics on these distilled models in chronological order, creating distilled metric sequence. We finally integrate distilled multi-metric sequences as a sequential multiformat and employ an attention-based RNN attack model for inference. Empirical results show SeqMIA outperforms all baselines, especially can achieve an order of magnitude improvement in terms of TPR @ 0.1% FPR. Furthermore, we delve into the reasons why this signal contributes to SeqMIA's high attack performance, and assess various defense mechanisms against SeqMIA.
Continual learning (CL) breaks off the one-way training manner and enables a model to adapt to new data, semantics and tasks continuously. However, current CL methods mainly focus on single tasks. Besides, CL models are plagued by catastrophic forgetting and semantic drift since the lack of old data, which often occurs in remote-sensing interpretation due to the intricate fine-grained semantics. In this paper, we propose Continual Panoptic Perception (CPP), a unified continual learning model that leverages multi-task joint learning covering pixel-level classification, instance-level segmentation and image-level perception for universal interpretation in remote sensing images. Concretely, we propose a collaborative cross-modal encoder (CCE) to extract the input image features, which supports pixel classification and caption generation synchronously. To inherit the knowledge from the old model without exemplar memory, we propose a task-interactive knowledge distillation (TKD) method, which leverages cross-modal optimization and task-asymmetric pseudo-labeling (TPL) to alleviate catastrophic forgetting. Furthermore, we also propose a joint optimization mechanism to achieve end-to-end multi-modal panoptic perception. Experimental results on the fine-grained panoptic perception dataset validate the effectiveness of the proposed model, and also prove that joint optimization can boost sub-task CL efficiency with over 13\% relative improvement on panoptic quality.
The universality of deep neural networks across different modalities and their generalization capabilities to unseen domains play an essential role in medical image segmentation. The recent Segment Anything Model (SAM) has demonstrated its potential in both settings. However, the huge computational costs, demand for manual annotations as prompts and conflict-prone decoding process of SAM degrade its generalizability and applicability in clinical scenarios. To address these issues, we propose an efficient self-prompting SAM for universal domain-generalized medical image segmentation, named ESP-MedSAM. Specifically, we first devise the Multi-Modal Decoupled Knowledge Distillation (MMDKD) strategy to construct a lightweight semi-parameter sharing image encoder that produces discriminative visual features for diverse modalities. Further, we introduce the Self-Patch Prompt Generator (SPPG) to automatically generate high-quality dense prompt embeddings for guiding segmentation decoding. Finally, we design the Query-Decoupled Modality Decoder (QDMD) that leverages a one-to-one strategy to provide an independent decoding channel for every modality. Extensive experiments indicate that ESP-MedSAM outperforms state-of-the-arts in diverse medical imaging segmentation tasks, displaying superior modality universality and generalization capabilities. Especially, ESP-MedSAM uses only 4.5\% parameters compared to SAM-H. The source code is available at https://github.com/xq141839/ESP-MedSAM.
Despite its breakthrough in classification problems, Knowledge distillation (KD) to recommendation models and ranking problems has not been studied well in the previous literature. This dissertation is devoted to developing knowledge distillation methods for recommender systems to fully improve the performance of a compact model. We propose novel distillation methods designed for recommender systems. The proposed methods are categorized according to their knowledge sources as follows: (1) Latent knowledge: we propose two methods that transfer latent knowledge of user/item representation. They effectively transfer knowledge of niche tastes with a balanced distillation strategy that prevents the KD process from being biased towards a small number of large preference groups. Also, we propose a new method that transfers user/item relations in the representation space. The proposed method selectively transfers essential relations considering the limited capacity of the compact model. (2) Ranking knowledge: we propose three methods that transfer ranking knowledge from the recommendation results. They formulate the KD process as a ranking matching problem and transfer the knowledge via a listwise learning strategy. Further, we present a new learning framework that compresses the ranking knowledge of heterogeneous recommendation models. The proposed framework is developed to ease the computational burdens of model ensemble which is a dominant solution for many recommendation applications. We validate the benefit of our proposed methods and frameworks through extensive experiments. To summarize, this dissertation sheds light on knowledge distillation approaches for a better accuracy-efficiency trade-off of the recommendation models.
We introduce the problem of continual distillation learning (CDL) in order to use knowledge distillation (KD) to improve prompt-based continual learning (CL) models. The CDL problem is valuable to study since the use of a larger vision transformer (ViT) leads to better performance in prompt-based continual learning. The distillation of knowledge from a large ViT to a small ViT improves the inference efficiency for prompt-based CL models. We empirically found that existing KD methods such as logit distillation and feature distillation cannot effectively improve the student model in the CDL setup. To address this issue, we introduce a novel method named Knowledge Distillation based on Prompts (KDP), in which globally accessible prompts specifically designed for knowledge distillation are inserted into the frozen ViT backbone of the student model. We demonstrate that our KDP method effectively enhances the distillation performance in comparison to existing KD methods in the CDL setup.
This paper discusses four facets of the Knowledge Distillation (KD) process for Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) architectures, particularly when executed on edge devices with constrained processing capabilities. First, we conduct a comparative analysis of the KD process between CNNs and ViT architectures, aiming to elucidate the feasibility and efficacy of employing different architectural configurations for the teacher and student, while assessing their performance and efficiency. Second, we explore the impact of varying the size of the student model on accuracy and inference speed, while maintaining a constant KD duration. Third, we examine the effects of employing higher resolution images on the accuracy, memory footprint and computational workload. Last, we examine the performance improvements obtained by fine-tuning the student model after KD to specific downstream tasks. Through empirical evaluations and analyses, this research provides AI practitioners with insights into optimal strategies for maximizing the effectiveness of the KD process on edge devices.
Large language models (LLMs) have shown remarkable advancements in chemistry and biomedical research, acting as versatile foundation models for various tasks. We introduce AMP-Designer, an LLM-based approach for swiftly designing novel antimicrobial peptides (AMPs) with desired properties. Within 11 days, AMP-Designer achieved the de novo design of 18 AMPs with broad-spectrum activity against Gram-negative bacteria. In vitro validation revealed a 94.4% success rate, with two candidates demonstrating exceptional antibacterial efficacy, minimal hemotoxicity, stability in human plasma, and low potential to induce resistance, as evidenced by significant bacterial load reduction in murine lung infection experiments. The entire process, from design to validation, concluded in 48 days. AMP-Designer excels in creating AMPs targeting specific strains despite limited data availability, with a top candidate displaying a minimum inhibitory concentration of 2.0 {\mu}g/ml against Propionibacterium acnes. Integrating advanced machine learning techniques, AMP-Designer demonstrates remarkable efficiency, paving the way for innovative solutions to antibiotic resistance.
In recent years, current mainstream feature masking distillation methods mainly function by reconstructing selectively masked regions of a student network from the feature maps of a teacher network. In these methods, attention mechanisms can help to identify spatially important regions and crucial object-aware channel clues, such that the reconstructed features are encoded with sufficient discriminative and representational power similar to teacher features. However, previous feature-masking distillation methods mainly address homogeneous knowledge distillation without fully taking into account the heterogeneous knowledge distillation scenario. In particular, the huge discrepancy between the teacher and the student frameworks within the heterogeneous distillation paradigm is detrimental to feature masking, leading to deteriorating reconstructed student features. In this study, a novel dual feature-masking heterogeneous distillation framework termed DFMSD is proposed for object detection. More specifically, a stage-wise adaptation learning module is incorporated into the dual feature-masking framework, and thus the student model can be progressively adapted to the teacher models for bridging the gap between heterogeneous networks. Furthermore, a masking enhancement strategy is combined with stage-wise learning such that object-aware masking regions are adaptively strengthened to improve feature-masking reconstruction. In addition, semantic alignment is performed at each Feature Pyramid Network (FPN) layer between the teacher and the student networks for generating consistent feature distributions. Our experiments for the object detection task demonstrate the promise of our approach, suggesting that DFMSD outperforms both the state-of-the-art heterogeneous and homogeneous distillation methods.
The scarcity of large-scale 3D-text paired data poses a great challenge on open vocabulary 3D scene understanding, and hence it is popular to leverage internet-scale 2D data and transfer their open vocabulary capabilities to 3D models through knowledge distillation. However, the existing distillation-based 3D scene understanding approaches rely on the representation capacity of 2D models, disregarding the exploration of geometric priors and inherent representational advantages offered by 3D data. In this paper, we propose an effective approach, namely Geometry Guided Self-Distillation (GGSD), to learn superior 3D representations from 2D pre-trained models. Specifically, we first design a geometry guided distillation module to distill knowledge from 2D models, and then leverage the 3D geometric priors to alleviate the inherent noise in 2D models and enhance the representation learning process. Due to the advantages of 3D representation, the performance of the distilled 3D student model can significantly surpass that of the 2D teacher model. This motivates us to further leverage the representation advantages of 3D data through self-distillation. As a result, our proposed GGSD approach outperforms the existing open vocabulary 3D scene understanding methods by a large margin, as demonstrated by our experiments on both indoor and outdoor benchmark datasets.