PyTorchKR Papers

최신 논문

총 1,512편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2403.09919공식 코드 · Papers with Code

Recurrent Drafter for Fast Speculative Decoding in Large Language Models

We present Recurrent Drafter (ReDrafter), an advanced speculative decoding approach that achieves state-of-the-art speedup for large language models (LLMs) inference. The performance gains are driven by three key aspects: (1) leveraging a recurrent neural network (RNN) as the draft model conditioning on LLM's hidden states, (2) applying a dynamic tree attention algorithm over beam search results to eliminate duplicated prefixes in candidate sequences, and (3) training through knowledge distillation from the LLM. ReDrafter accelerates Vicuna inference in MT-Bench by up to 2.8x with a PyTorch implementation on Nvidia H100 GPUs. To demonstrate its practicality in real environments, we also validated its effectiveness for on-device applications by implementing the approach in MLX and benchmarking performance on Metal GPUs in Apple Silicon chips, achieving up to 2.3x speedup.

arXiv:2403.09522공식 코드 · Papers with Code

MT-PATCHER: Selective and Extendable Knowledge Distillation from Large Language Models for Machine Translation

Large Language Models (LLM) have demonstrated their strong ability in the field of machine translation (MT), yet they suffer from high computational cost and latency. Therefore, transferring translation knowledge from giant LLMs to medium-sized machine translation models is a promising research direction. However, traditional knowledge distillation methods do not take the capability of student and teacher models into consideration, therefore repeatedly teaching student models on the knowledge they have learned, and failing to extend to novel contexts and knowledge. In this paper, we propose a framework called MT-Patcher, which transfers knowledge from LLMs to existing MT models in a selective, comprehensive and proactive manner. Considering the current translation ability of student MT models, we only identify and correct their translation errors, instead of distilling the whole translation from the teacher. Leveraging the strong language abilities of LLMs, we instruct LLM teachers to synthesize diverse contexts and anticipate more potential errors for the student. Experiment results on translating both specific language phenomena and general MT benchmarks demonstrate that finetuning the student MT model on about 10% examples can achieve comparable results to the traditional knowledge distillation method, and synthesized potential errors and diverse contexts further improve translation performances on unseen contexts and words.

arXiv:2403.09486공식 코드 · Papers with Code

SpikeReveal: Unlocking Temporal Sequences from Real Blurry Inputs with Spike Streams

Reconstructing a sequence of sharp images from the blurry input is crucial for enhancing our insights into the captured scene and poses a significant challenge due to the limited temporal features embedded in the image. Spike cameras, sampling at rates up to 40,000 Hz, have proven effective in capturing motion features and beneficial for solving this ill-posed problem. Nonetheless, existing methods fall into the supervised learning paradigm, which suffers from notable performance degradation when applied to real-world scenarios that diverge from the synthetic training data domain. Moreover, the quality of reconstructed images is capped by the generated images based on motion analysis interpolation, which inherently differs from the actual scene, affecting the generalization ability of these methods in real high-speed scenarios. To address these challenges, we propose the first self-supervised framework for the task of spike-guided motion deblurring. Our approach begins with the formulation of a spike-guided deblurring model that explores the theoretical relationships among spike streams, blurry images, and their corresponding sharp sequences. We subsequently develop a self-supervised cascaded framework to alleviate the issues of spike noise and spatial-resolution mismatching encountered in the deblurring model. With knowledge distillation and re-blurring loss, we further design a lightweight deblur network to generate high-quality sequences with brightness and texture consistency with the original input. Quantitative and qualitative experiments conducted on our real-world and synthetic datasets with spikes validate the superior generalization of the proposed framework. Our code, data and trained models will be available at \url{https://github.com/chenkang455/S-SDM}.

arXiv:2403.09313공식 코드 · Papers with Code

Knowledge Distillation in YOLOX-ViT for Side-Scan Sonar Object Detection

In this paper we present YOLOX-ViT, a novel object detection model, and investigate the efficacy of knowledge distillation for model size reduction without sacrificing performance. Focused on underwater robotics, our research addresses key questions about the viability of smaller models and the impact of the visual transformer layer in YOLOX. Furthermore, we introduce a new side-scan sonar image dataset, and use it to evaluate our object detector's performance. Results show that knowledge distillation effectively reduces false positives in wall detection. Additionally, the introduced visual transformer layer significantly improves object detection accuracy in the underwater environment. The source code of the knowledge distillation in the YOLOX-ViT is at https://github.com/remaro-network/KD-YOLOX-ViT.

arXiv:2403.08654공식 코드 · Papers with Code

An Efficient End-to-End Approach to Noise Invariant Speech Features via Multi-Task Learning

Self-supervised speech representation learning enables the extraction of meaningful features from raw waveforms. These features can then be efficiently used across multiple downstream tasks. However, two significant issues arise when considering the deployment of such methods ``in-the-wild": (i) Their large size, which can be prohibitive for edge applications; and (ii) their robustness to detrimental factors, such as noise and/or reverberation, that can heavily degrade the performance of such systems. In this work, we propose RobustDistiller, a novel knowledge distillation mechanism that tackles both problems jointly. Simultaneously to the distillation recipe, we apply a multi-task learning objective to encourage the network to learn noise-invariant representations by denoising the input. The proposed mechanism is evaluated on twelve different downstream tasks. It outperforms several benchmarks regardless of noise type, or noise and reverberation levels. Experimental results show that the new Student model with 23M parameters can achieve results comparable to the Teacher model with 95M parameters. Lastly, we show that the proposed recipe can be applied to other distillation methodologies, such as the recent DPWavLM. For reproducibility, code and model checkpoints will be made available at \mbox{\url{https://github.com/Hguimaraes/robustdistiller}}.

arXiv:2403.08040공식 코드 · Papers with Code

Low-Energy On-Device Personalization for MCUs

Microcontroller Units (MCUs) are ideal platforms for edge applications due to their low cost and energy consumption, and are widely used in various applications, including personalized machine learning tasks, where customized models can enhance the task adaptation. However, existing approaches for local on-device personalization mostly support simple ML architectures or require complex local pre-training/training, leading to high energy consumption and negating the low-energy advantage of MCUs. In this paper, we introduce $MicroT$, an efficient and low-energy MCU personalization approach. $MicroT$ includes a robust, general, but tiny feature extractor, developed through self-supervised knowledge distillation, which trains a task-specific head to enable independent on-device personalization with minimal energy and computational requirements. MicroT implements an MCU-optimized early-exit inference mechanism called stage-decision to further reduce energy costs. This mechanism allows for user-configurable exit criteria (stage-decision ratio) to adaptively balance energy cost with model performance. We evaluated MicroT using two models, three datasets, and two MCU boards. $MicroT$ outperforms traditional transfer learning (TTL) and two SOTA approaches by 2.12 - 11.60% across two models and three datasets. Targeting widely used energy-aware edge devices, MicroT's on-device training requires no additional complex operations, halving the energy cost compared to SOTA approaches by up to 2.28X while keeping SRAM usage below 1MB. During local inference, MicroT reduces energy cost by 14.17% compared to TTL across two boards and two datasets, highlighting its suitability for long-term use on energy-aware resource-constrained MCUs.

arXiv:2403.07300공식 코드 · Papers with Code

CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning

Deep learning (e.g., Transformer) has been widely and successfully used in multivariate time series forecasting (MTSF). Unlike existing methods that focus on training models from a single modal of time series input, large language models (LLMs) based MTSF methods with cross-modal text and time series input have recently shown great superiority, especially with limited temporal data. However, current LLM-based MTSF methods usually focus on adapting and fine-tuning LLMs, while neglecting the distribution discrepancy between textual and temporal input tokens, thus leading to sub-optimal performance. To address this issue, we propose a novel Cross-Modal LLM Fine-Tuning (CALF) framework for MTSF by reducing the distribution discrepancy between textual and temporal data, which mainly consists of the temporal target branch with temporal input and the textual source branch with aligned textual input. To reduce the distribution discrepancy, we develop the cross-modal match module to first align cross-modal input distributions. Additionally, to minimize the modality distribution gap in both feature and output spaces, feature regularization loss is developed to align the intermediate features between the two branches for better weight updates, while output consistency loss is introduced to allow the output representations of both branches to correspond effectively. Thanks to the modality alignment, CALF establishes state-of-the-art performance for both long-term and short-term forecasting tasks with low computational complexity, and exhibiting favorable few-shot and zero-shot abilities similar to that in LLMs. Code is available at \url{https://github.com/Hank0626/LLaTA}.

arXiv:2403.07030공식 코드 · Papers with Code

AuG-KD: Anchor-Based Mixup Generation for Out-of-Domain Knowledge Distillation

Due to privacy or patent concerns, a growing number of large models are released without granting access to their training data, making transferring their knowledge inefficient and problematic. In response, Data-Free Knowledge Distillation (DFKD) methods have emerged as direct solutions. However, simply adopting models derived from DFKD for real-world applications suffers significant performance degradation, due to the discrepancy between teachers' training data and real-world scenarios (student domain). The degradation stems from the portions of teachers' knowledge that are not applicable to the student domain. They are specific to the teacher domain and would undermine students' performance. Hence, selectively transferring teachers' appropriate knowledge becomes the primary challenge in DFKD. In this work, we propose a simple but effective method AuG-KD. It utilizes an uncertainty-guided and sample-specific anchor to align student-domain data with the teacher domain and leverages a generative method to progressively trade off the learning process between OOD knowledge distillation and domain-specific information learning via mixup learning. Extensive experiments in 3 datasets and 8 settings demonstrate the stability and superiority of our approach. Code available at https://github.com/IshiKura-a/AuG-KD .

arXiv:2403.07292공식 코드 · Papers with Code

Continual All-in-One Adverse Weather Removal with Knowledge Replay on a Unified Network Structure

In real-world applications, image degeneration caused by adverse weather is always complex and changes with different weather conditions from days and seasons. Systems in real-world environments constantly encounter adverse weather conditions that are not previously observed. Therefore, it practically requires adverse weather removal models to continually learn from incrementally collected data reflecting various degeneration types. Existing adverse weather removal approaches, for either single or multiple adverse weathers, are mainly designed for a static learning paradigm, which assumes that the data of all types of degenerations to handle can be finely collected at one time before a single-phase learning process. They thus cannot directly handle the incremental learning requirements. To address this issue, we made the earliest effort to investigate the continual all-in-one adverse weather removal task, in a setting closer to real-world applications. Specifically, we develop a novel continual learning framework with effective knowledge replay (KR) on a unified network structure. Equipped with a principal component projection and an effective knowledge distillation mechanism, the proposed KR techniques are tailored for the all-in-one weather removal task. It considers the characteristics of the image restoration task with multiple degenerations in continual learning, and the knowledge for different degenerations can be shared and accumulated in the unified network structure. Extensive experimental results demonstrate the effectiveness of the proposed method to deal with this challenging task, which performs competitively to existing dedicated or joint training image restoration methods. Our code is available at https://github.com/xiaojihh/CL_all-in-one.

arXiv:2403.06071공식 코드 · Papers with Code

Bit-mask Robust Contrastive Knowledge Distillation for Unsupervised Semantic Hashing

Unsupervised semantic hashing has emerged as an indispensable technique for fast image search, which aims to convert images into binary hash codes without relying on labels. Recent advancements in the field demonstrate that employing large-scale backbones (e.g., ViT) in unsupervised semantic hashing models can yield substantial improvements. However, the inference delay has become increasingly difficult to overlook. Knowledge distillation provides a means for practical model compression to alleviate this delay. Nevertheless, the prevailing knowledge distillation approaches are not explicitly designed for semantic hashing. They ignore the unique search paradigm of semantic hashing, the inherent necessities of the distillation process, and the property of hash codes. In this paper, we propose an innovative Bit-mask Robust Contrastive knowledge Distillation (BRCD) method, specifically devised for the distillation of semantic hashing models. To ensure the effectiveness of two kinds of search paradigms in the context of semantic hashing, BRCD first aligns the semantic spaces between the teacher and student models through a contrastive knowledge distillation objective. Additionally, to eliminate noisy augmentations and ensure robust optimization, a cluster-based method within the knowledge distillation process is introduced. Furthermore, through a bit-level analysis, we uncover the presence of redundancy bits resulting from the bit independence property. To mitigate these effects, we introduce a bit mask mechanism in our knowledge distillation objective. Finally, extensive experiments not only showcase the noteworthy performance of our BRCD method in comparison to other knowledge distillation methods but also substantiate the generality of our methods across diverse semantic hashing models and backbones. The code for BRCD is available at https://github.com/hly1998/BRCD.

arXiv:2403.05894공식 코드 · Papers with Code

Frequency Attention for Knowledge Distillation

Knowledge distillation is an attractive approach for learning compact deep neural networks, which learns a lightweight student model by distilling knowledge from a complex teacher model. Attention-based knowledge distillation is a specific form of intermediate feature-based knowledge distillation that uses attention mechanisms to encourage the student to better mimic the teacher. However, most of the previous attention-based distillation approaches perform attention in the spatial domain, which primarily affects local regions in the input image. This may not be sufficient when we need to capture the broader context or global information necessary for effective knowledge transfer. In frequency domain, since each frequency is determined from all pixels of the image in spatial domain, it can contain global information about the image. Inspired by the benefits of the frequency domain, we propose a novel module that functions as an attention mechanism in the frequency domain. The module consists of a learnable global filter that can adjust the frequencies of student's features under the guidance of the teacher's features, which encourages the student's features to have patterns similar to the teacher's features. We then propose an enhanced knowledge review-based distillation model by leveraging the proposed frequency attention module. The extensive experiments with various teacher and student architectures on image classification and object detection benchmark datasets show that the proposed approach outperforms other knowledge distillation methods.

arXiv:2403.05796공식 코드 · Papers with Code

Weakly Supervised Change Detection via Knowledge Distillation and Multiscale Sigmoid Inference

Change detection, which aims to detect spatial changes from a pair of multi-temporal images due to natural or man-made causes, has been widely applied in remote sensing, disaster management, urban management, etc. Most existing change detection approaches, however, are fully supervised and require labor-intensive pixel-level labels. To address this, we develop a novel weakly supervised change detection technique via Knowledge Distillation and Multiscale Sigmoid Inference (KD-MSI) that leverages image-level labels. In our approach, the Class Activation Maps (CAM) are utilized not only to derive a change probability map but also to serve as a foundation for the knowledge distillation process. This is done through a joint training strategy of the teacher and student networks, enabling the student network to highlight potential change areas more accurately than teacher network based on image-level labels. Moreover, we designed a Multiscale Sigmoid Inference (MSI) module as a post processing step to further refine the change probability map from the trained student network. Empirical results on three public datasets, i.e., WHU-CD, DSIFN-CD, and LEVIR-CD, demonstrate that our proposed technique, with its integrated training strategy, significantly outperforms the state-of-the-art.

arXiv:2403.06679공식 코드 · Papers with Code

Answering Diverse Questions via Text Attached with Key Audio-Visual Clues

Audio-visual question answering (AVQA) requires reference to video content and auditory information, followed by correlating the question to predict the most precise answer. Although mining deeper layers of audio-visual information to interact with questions facilitates the multimodal fusion process, the redundancy of audio-visual parameters tends to reduce the generalization of the inference engine to multiple question-answer pairs in a single video. Indeed, the natural heterogeneous relationship between audiovisuals and text makes the perfect fusion challenging, to prevent high-level audio-visual semantics from weakening the network's adaptability to diverse question types, we propose a framework for performing mutual correlation distillation (MCD) to aid question inference. MCD is divided into three main steps: 1) firstly, the residual structure is utilized to enhance the audio-visual soft associations based on self-attention, then key local audio-visual features relevant to the question context are captured hierarchically by shared aggregators and coupled in the form of clues with specific question vectors. 2) Secondly, knowledge distillation is enforced to align audio-visual-text pairs in a shared latent space to narrow the cross-modal semantic gap. 3) And finally, the audio-visual dependencies are decoupled by discarding the decision-level integrations. We evaluate the proposed method on two publicly available datasets containing multiple question-and-answer pairs, i.e., Music-AVQA and AVQA. Experiments show that our method outperforms other state-of-the-art methods, and one interesting finding behind is that removing deep audio-visual features during inference can effectively mitigate overfitting. The source code is released at http://github.com/rikeilong/MCD-forAVQA.

arXiv:2403.06239공식 코드 · Papers with Code

Cooperative Classification and Rationalization for Graph Generalization

Graph Neural Networks (GNNs) have achieved impressive results in graph classification tasks, but they struggle to generalize effectively when faced with out-of-distribution (OOD) data. Several approaches have been proposed to address this problem. Among them, one solution is to diversify training distributions in vanilla classification by modifying the data environment, yet accessing the environment information is complex. Besides, another promising approach involves rationalization, extracting invariant rationales for predictions. However, extracting rationales is difficult due to limited learning signals, resulting in less accurate rationales and diminished predictions. To address these challenges, in this paper, we propose a Cooperative Classification and Rationalization (C2R) method, consisting of the classification and the rationalization module. Specifically, we first assume that multiple environments are available in the classification module. Then, we introduce diverse training distributions using an environment-conditional generative network, enabling robust graph representations. Meanwhile, the rationalization module employs a separator to identify relevant rationale subgraphs while the remaining non-rationale subgraphs are de-correlated with labels. Next, we align graph representations from the classification module with rationale subgraph representations using the knowledge distillation methods, enhancing the learning signal for rationales. Finally, we infer multiple environments by gathering non-rationale representations and incorporate them into the classification module for cooperative learning. Extensive experimental results on both benchmarks and synthetic datasets demonstrate the effectiveness of C2R. Code is available at https://github.com/yuelinan/Codes-of-C2R.

arXiv:2403.06213공식 코드 · Papers with Code

$V_kD:$ Improving Knowledge Distillation using Orthogonal Projections

Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a 4.4% relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available: https://github.com/roymiles/vkd

arXiv:2403.06914공식 코드 · Papers with Code

MEND: Meta dEmonstratioN Distillation for Efficient and Effective In-Context Learning

Large Language models (LLMs) have demonstrated impressive in-context learning (ICL) capabilities, where a LLM makes predictions for a given test input together with a few input-output pairs (demonstrations). Nevertheless, the inclusion of demonstrations leads to a quadratic increase in the computational overhead of the self-attention mechanism. Existing solutions attempt to distill lengthy demonstrations into compact vectors. However, they often require task-specific retraining or compromise LLM's in-context learning performance. To mitigate these challenges, we present Meta dEmonstratioN Distillation (MEND), where a language model learns to distill any lengthy demonstrations into vectors without retraining for a new downstream task. We exploit the knowledge distillation to enhance alignment between MEND and LLM, achieving both efficiency and effectiveness simultaneously. MEND is endowed with the meta-knowledge of distilling demonstrations through a two-stage training process, which includes meta-distillation pretraining and fine-tuning. Comprehensive evaluations across seven diverse ICL task partitions using decoder-only (GPT-2) and encoder-decoder (T5) attest to MEND's prowess. It not only matches but often outperforms the Vanilla ICL as well as other state-of-the-art distillation models, while significantly reducing the computational demands. This innovation promises enhanced scalability and efficiency for the practical deployment of large language models

arXiv:2403.05061공식 코드 · Papers with Code

RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features

The inherent noisy and sparse characteristics of radar data pose challenges in finding effective representations for 3D object detection. In this paper, we propose RadarDistill, a novel knowledge distillation (KD) method, which can improve the representation of radar data by leveraging LiDAR data. RadarDistill successfully transfers desirable characteristics of LiDAR features into radar features using three key components: Cross-Modality Alignment (CMA), Activation-based Feature Distillation (AFD), and Proposal-based Feature Distillation (PFD). CMA enhances the density of radar features by employing multiple layers of dilation operations, effectively addressing the challenge of inefficient knowledge transfer from LiDAR to radar. AFD selectively transfers knowledge based on regions of the LiDAR features, with a specific focus on areas where activation intensity exceeds a predefined threshold. PFD similarly guides the radar network to selectively mimic features from the LiDAR network within the object proposals. Our comparative analyses conducted on the nuScenes datasets demonstrate that RadarDistill achieves state-of-the-art (SOTA) performance for radar-only object detection task, recording 20.5% in mAP and 43.7% in NDS. Also, RadarDistill significantly improves the performance of the camera-radar fusion model.

arXiv:2403.04908공식 코드 · Papers with Code

Self-Adapting Large Visual-Language Models to Edge Devices across Visual Modalities

Recent advancements in Vision-Language (VL) models have sparked interest in their deployment on edge devices, yet challenges in handling diverse visual modalities, manual annotation, and computational constraints remain. We introduce EdgeVL, a novel framework that bridges this gap by seamlessly integrating dual-modality knowledge distillation and quantization-aware contrastive learning. This approach enables the adaptation of large VL models, like CLIP, for efficient use with both RGB and non-RGB images on resource-limited devices without the need for manual annotations. EdgeVL not only transfers visual language alignment capabilities to compact models but also maintains feature quality post-quantization, significantly enhancing open-vocabulary classification performance across various visual modalities. Our work represents the first systematic effort to adapt large VL models for edge deployment, showcasing up to 15.4% accuracy improvements on multiple datasets and up to 93-fold reduction in model size.

arXiv:2403.04245공식 코드 · Papers with Code

A Study of Dropout-Induced Modality Bias on Robustness to Missing Video Frames for Audio-Visual Speech Recognition

Advanced Audio-Visual Speech Recognition (AVSR) systems have been observed to be sensitive to missing video frames, performing even worse than single-modality models. While applying the dropout technique to the video modality enhances robustness to missing frames, it simultaneously results in a performance loss when dealing with complete data input. In this paper, we investigate this contrasting phenomenon from the perspective of modality bias and reveal that an excessive modality bias on the audio caused by dropout is the underlying reason. Moreover, we present the Modality Bias Hypothesis (MBH) to systematically describe the relationship between modality bias and robustness against missing modality in multimodal systems. Building on these findings, we propose a novel Multimodal Distribution Approximation with Knowledge Distillation (MDA-KD) framework to reduce over-reliance on the audio modality and to maintain performance and robustness simultaneously. Finally, to address an entirely missing modality, we adopt adapters to dynamically switch decision strategies. The effectiveness of our proposed approach is evaluated and validated through a series of comprehensive experiments using the MISP2021 and MISP2022 datasets. Our code is available at https://github.com/dalision/ModalBiasAVSR

arXiv:2403.03483공식 코드 · Papers with Code

A Teacher-Free Graph Knowledge Distillation Framework with Dual Self-Distillation

Recent years have witnessed great success in handling graph-related tasks with Graph Neural Networks (GNNs). Despite their great academic success, Multi-Layer Perceptrons (MLPs) remain the primary workhorse for practical industrial applications. One reason for such an academic-industry gap is the neighborhood-fetching latency incurred by data dependency in GNNs. To reduce their gaps, Graph Knowledge Distillation (GKD) is proposed, usually based on a standard teacher-student architecture, to distill knowledge from a large teacher GNN into a lightweight student GNN or MLP. However, we found in this paper that neither teachers nor GNNs are necessary for graph knowledge distillation. We propose a Teacher-Free Graph Self-Distillation (TGS) framework that does not require any teacher model or GNNs during both training and inference. More importantly, the proposed TGS framework is purely based on MLPs, where structural information is only implicitly used to guide dual knowledge self-distillation between the target node and its neighborhood. As a result, TGS enjoys the benefits of graph topology awareness in training but is free from data dependency in inference. Extensive experiments have shown that the performance of vanilla MLPs can be greatly improved with dual self-distillation, e.g., TGS improves over vanilla MLPs by 15.54% on average and outperforms state-of-the-art GKD algorithms on six real-world datasets. In terms of inference speed, TGS infers 75X-89X faster than existing GNNs and 16X-25X faster than classical inference acceleration methods.