PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2401.06356

An Empirical Investigation into the Effect of Parameter Choices in Knowledge Distillation

We present a large-scale empirical study of how choices of configuration parameters affect performance in knowledge distillation (KD). An example of such a KD parameter is the measure of distance between the predictions of the teacher and the student, common choices for which include the mean squared error (MSE) and the KL-divergence. Although scattered efforts have been made to understand the differences between such options, the KD literature still lacks a systematic study on their general effect on student performance. We take an empirical approach to this question in this paper, seeking to find out the extent to which such choices influence student performance across 13 datasets from 4 NLP tasks and 3 student sizes. We quantify the cost of making sub-optimal choices and identify a single configuration that performs well across the board.

arXiv:2401.06010공식 코드 · Papers with Code

Attention to detail: inter-resolution knowledge distillation

The development of computer vision solutions for gigapixel images in digital pathology is hampered by significant computational limitations due to the large size of whole slide images. In particular, digitizing biopsies at high resolutions is a time-consuming process, which is necessary due to the worsening results from the decrease in image detail. To alleviate this issue, recent literature has proposed using knowledge distillation to enhance the model performance at reduced image resolutions. In particular, soft labels and features extracted at the highest magnification level are distilled into a model that takes lower-magnification images as input. However, this approach fails to transfer knowledge about the most discriminative image regions in the classification process, which may be lost when the resolution is decreased. In this work, we propose to distill this information by incorporating attention maps during training. In particular, our formulation leverages saliency maps of the target class via grad-CAMs, which guides the lower-resolution Student model to match the Teacher distribution by minimizing the l2 distance between them. Comprehensive experiments on prostate histology image grading demonstrate that the proposed approach substantially improves the model performance across different image resolutions compared to previous literature.

arXiv:2401.05735

Object-Centric Diffusion for Efficient Video Editing

Diffusion-based video editing have reached impressive quality and can transform either the global style, local structure, and attributes of given video inputs, following textual edit prompts. However, such solutions typically incur heavy memory and computational costs to generate temporally-coherent frames, either in the form of diffusion inversion and/or cross-frame attention. In this paper, we conduct an analysis of such inefficiencies, and suggest simple yet effective modifications that allow significant speed-ups whilst maintaining quality. Moreover, we introduce Object-Centric Diffusion, to fix generation artifacts and further reduce latency by allocating more computations towards foreground edited regions, arguably more important for perceptual quality. We achieve this by two novel proposals: i) Object-Centric Sampling, decoupling the diffusion steps spent on salient or background regions and spending most on the former, and ii) Object-Centric Token Merging, which reduces cost of cross-frame attention by fusing redundant tokens in unimportant background regions. Both techniques are readily applicable to a given video editing model without retraining, and can drastically reduce its memory and computational cost. We evaluate our proposals on inversion-based and control-signal-based editing pipelines, and show a latency reduction up to 10x for a comparable synthesis quality. Project page: qualcomm-ai-research.github.io/object-centric-diffusion.

arXiv:2401.05676

Exploring Self- and Cross-Triplet Correlations for Human-Object Interaction Detection

Human-Object Interaction (HOI) detection plays a vital role in scene understanding, which aims to predict the HOI triplet in the form of <human, object, action>. Existing methods mainly extract multi-modal features (e.g., appearance, object semantics, human pose) and then fuse them together to directly predict HOI triplets. However, most of these methods focus on seeking for self-triplet aggregation, but ignore the potential cross-triplet dependencies, resulting in ambiguity of action prediction. In this work, we propose to explore Self- and Cross-Triplet Correlations (SCTC) for HOI detection. Specifically, we regard each triplet proposal as a graph where Human, Object represent nodes and Action indicates edge, to aggregate self-triplet correlation. Also, we try to explore cross-triplet dependencies by jointly considering instance-level, semantic-level, and layout-level relations. Besides, we leverage the CLIP model to assist our SCTC obtain interaction-aware feature by knowledge distillation, which provides useful action clues for HOI detection. Extensive experiments on HICO-DET and V-COCO datasets verify the effectiveness of our proposed SCTC.

arXiv:2401.04810공식 코드 · Papers with Code

Translate-Distill: Learning Cross-Language Dense Retrieval by Translation and Distillation

Prior work on English monolingual retrieval has shown that a cross-encoder trained using a large number of relevance judgments for query-document pairs can be used as a teacher to train more efficient, but similarly effective, dual-encoder student models. Applying a similar knowledge distillation approach to training an efficient dual-encoder model for Cross-Language Information Retrieval (CLIR), where queries and documents are in different languages, is challenging due to the lack of a sufficiently large training collection when the query and document languages differ. The state of the art for CLIR thus relies on translating queries, documents, or both from the large English MS MARCO training set, an approach called Translate-Train. This paper proposes an alternative, Translate-Distill, in which knowledge distillation from either a monolingual cross-encoder or a CLIR cross-encoder is used to train a dual-encoder CLIR student model. This richer design space enables the teacher model to perform inference in an optimized setting, while training the student model directly for CLIR. Trained models and artifacts are publicly available on Huggingface.

arXiv:2401.03430

Multi-Channel Multi-Domain based Knowledge Distillation Algorithm for Sleep Staging with Single-Channel EEG

This paper proposed a Multi-Channel Multi-Domain (MCMD) based knowledge distillation algorithm for sleep staging using single-channel EEG. Both knowledge from different domains and different channels are learnt in the proposed algorithm, simultaneously. A multi-channel pre-training and single-channel fine-tuning scheme is used in the proposed work. The knowledge from different channels in the source domain is transferred to the single-channel model in the target domain. A pre-trained teacher-student model scheme is used to distill knowledge from the multi-channel teacher model to the single-channel student model combining with output transfer and intermediate feature transfer in the target domain. The proposed algorithm achieves a state-of-the-art single-channel sleep staging accuracy of 86.5%, with only 0.6% deterioration from the state-of-the-art multi-channel model. There is an improvement of 2% compared to the baseline model. The experimental results show that knowledge from multiple domains (different datasets) and multiple channels (e.g. EMG, EOG) could be transferred to single-channel sleep staging.

arXiv:2401.03685

Logits Poisoning Attack in Federated Distillation

Federated Distillation (FD) is a novel and promising distributed machine learning paradigm, where knowledge distillation is leveraged to facilitate a more efficient and flexible cross-device knowledge transfer in federated learning. By optimizing local models with knowledge distillation, FD circumvents the necessity of uploading large-scale model parameters to the central server, simultaneously preserving the raw data on local clients. Despite the growing popularity of FD, there is a noticeable gap in previous works concerning the exploration of poisoning attacks within this framework. This can lead to a scant understanding of the vulnerabilities to potential adversarial actions. To this end, we introduce FDLA, a poisoning attack method tailored for FD. FDLA manipulates logit communications in FD, aiming to significantly degrade model performance on clients through misleading the discrimination of private samples. Through extensive simulation experiments across a variety of datasets, attack scenarios, and FD configurations, we demonstrate that LPA effectively compromises client model accuracy, outperforming established baseline algorithms in this regard. Our findings underscore the critical need for robust defense mechanisms in FD settings to mitigate such adversarial threats.

arXiv:2401.03246공식 코드 · Papers with Code

SeqNAS: Neural Architecture Search for Event Sequence Classification

Neural Architecture Search (NAS) methods are widely used in various industries to obtain high quality taskspecific solutions with minimal human intervention. Event Sequences find widespread use in various industrial applications including churn prediction customer segmentation fraud detection and fault diagnosis among others. Such data consist of categorical and real-valued components with irregular timestamps. Despite the usefulness of NAS methods previous approaches only have been applied to other domains images texts or time series. Our work addresses this limitation by introducing a novel NAS algorithm SeqNAS specifically designed for event sequence classification. We develop a simple yet expressive search space that leverages commonly used building blocks for event sequence classification including multihead self attention convolutions and recurrent cells. To perform the search we adopt sequential Bayesian Optimization and utilize previously trained models as an ensemble of teachers to augment knowledge distillation. As a result of our work we demonstrate that our method surpasses state of the art NAS methods and popular architectures suitable for sequence classification and holds great potential for various industrial applications.

arXiv:2401.03000공식 코드 · Papers with Code

Bridging Modalities: Knowledge Distillation and Masked Training for Translating Multi-Modal Emotion Recognition to Uni-Modal, Speech-Only Emotion Recognition

This paper presents an innovative approach to address the challenges of translating multi-modal emotion recognition models to a more practical and resource-efficient uni-modal counterpart, specifically focusing on speech-only emotion recognition. Recognizing emotions from speech signals is a critical task with applications in human-computer interaction, affective computing, and mental health assessment. However, existing state-of-the-art models often rely on multi-modal inputs, incorporating information from multiple sources such as facial expressions and gestures, which may not be readily available or feasible in real-world scenarios. To tackle this issue, we propose a novel framework that leverages knowledge distillation and masked training techniques.

arXiv:2401.02677공식 코드 · Papers with Code

Progressive Knowledge Distillation Of Stable Diffusion XL Using Layer Level Loss

Stable Diffusion XL (SDXL) has become the best open source text-to-image model (T2I) for its versatility and top-notch image quality. Efficiently addressing the computational demands of SDXL models is crucial for wider reach and applicability. In this work, we introduce two scaled-down variants, Segmind Stable Diffusion (SSD-1B) and Segmind-Vega, with 1.3B and 0.74B parameter UNets, respectively, achieved through progressive removal using layer-level losses focusing on reducing the model size while preserving generative quality. We release these models weights at https://hf.co/Segmind. Our methodology involves the elimination of residual networks and transformer blocks from the U-Net structure of SDXL, resulting in significant reductions in parameters, and latency. Our compact models effectively emulate the original SDXL by capitalizing on transferred knowledge, achieving competitive results against larger multi-billion parameter SDXL. Our work underscores the efficacy of knowledge distillation coupled with layer-level losses in reducing model size while preserving the high-quality generative capabilities of SDXL, thus facilitating more accessible deployment in resource-constrained environments.

arXiv:2401.02046

CTC Blank Triggered Dynamic Layer-Skipping for Efficient CTC-based Speech Recognition

Deploying end-to-end speech recognition models with limited computing resources remains challenging, despite their impressive performance. Given the gradual increase in model size and the wide range of model applications, selectively executing model components for different inputs to improve the inference efficiency is of great interest. In this paper, we propose a dynamic layer-skipping method that leverages the CTC blank output from intermediate layers to trigger the skipping of the last few encoder layers for frames with high blank probabilities. Furthermore, we factorize the CTC output distribution and perform knowledge distillation on intermediate layers to reduce computation and improve recognition accuracy. Experimental results show that by utilizing the CTC blank, the encoder layer depth can be adjusted dynamically, resulting in 29% acceleration of the CTC model inference with minor performance degradation.

arXiv:2401.01473

Self-supervised Reflective Learning through Self-distillation and Online Clustering for Speaker Representation Learning

Speaker representation learning is crucial for voice recognition systems, with recent advances in self-supervised approaches reducing dependency on labeled data. Current two-stage iterative frameworks, while effective, suffer from significant computational overhead due to repeated rounds of clustering and training. They also struggle with noisy pseudo labels that can impair model learning. This paper introduces self-supervised reflective learning (SSRL), an improved framework that addresses these limitations by enabling continuous refinement of pseudo labels during training. Through a teacher-student architecture and online clustering mechanism, SSRL eliminates the need for iterative training rounds. To handle label noise, we incorporate noisy label modeling and pseudo label queues that maintain temporal consistency. Experiments on VoxCeleb show SSRL's superiority over current two-stage iterative approaches, surpassing the performance of a 5-round method in just a single training round. Ablation studies validate the contributions of key components like noisy label modeling and pseudo label queues. Moreover, consistent improvements in pseudo labeling and the convergence of cluster counts demonstrate SSRL's effectiveness in deciphering unlabeled data. This work marks an important advancement in efficient and accurate self-supervised speaker representation learning through the novel reflective learning paradigm.

arXiv:2401.02329공식 코드 · Papers with Code

Exploring Vacant Classes in Label-Skewed Federated Learning

Label skews, characterized by disparities in local label distribution across clients, pose a significant challenge in federated learning. As minority classes suffer from worse accuracy due to overfitting on local imbalanced data, prior methods often incorporate class-balanced learning techniques during local training. Although these methods improve the mean accuracy across all classes, we observe that vacant classes-referring to categories absent from a client's data distribution-remain poorly recognized. Besides, there is still a gap in the accuracy of local models on minority classes compared to the global model. This paper introduces FedVLS, a novel approach to label-skewed federated learning that integrates both vacant-class distillation and logit suppression simultaneously. Specifically, vacant-class distillation leverages knowledge distillation during local training on each client to retain essential information related to vacant classes from the global model. Moreover, logit suppression directly penalizes network logits for non-label classes, effectively addressing misclassifications in minority classes that may be biased toward majority classes. Extensive experiments validate the efficacy of FedVLS, demonstrating superior performance compared to previous state-of-the-art (SOTA) methods across diverse datasets with varying degrees of label skews. Our code is available at https://github.com/krumpguo/FedVLS.

arXiv:2401.01918

Distilling Temporal Knowledge with Masked Feature Reconstruction for 3D Object Detection

Striking a balance between precision and efficiency presents a prominent challenge in the bird's-eye-view (BEV) 3D object detection. Although previous camera-based BEV methods achieved remarkable performance by incorporating long-term temporal information, most of them still face the problem of low efficiency. One potential solution is knowledge distillation. Existing distillation methods only focus on reconstructing spatial features, while overlooking temporal knowledge. To this end, we propose TempDistiller, a Temporal knowledge Distiller, to acquire long-term memory from a teacher detector when provided with a limited number of frames. Specifically, a reconstruction target is formulated by integrating long-term temporal knowledge through self-attention operation applied to feature teachers. Subsequently, novel features are generated for masked student features via a generator. Ultimately, we utilize this reconstruction target to reconstruct the student features. In addition, we also explore temporal relational knowledge when inputting full frames for the student model. We verify the effectiveness of the proposed method on the nuScenes benchmark. The experimental results show our method obtain an enhancement of +1.6 mAP and +1.1 NDS compared to the baseline, a speed improvement of approximately 6 FPS after compressing temporal knowledge, and the most accurate velocity estimation.

arXiv:2401.02287공식 코드 · Papers with Code

Distillation-based fabric anomaly detection

Unsupervised texture anomaly detection has been a concerning topic in a vast amount of industrial processes. Patterned textures inspection, particularly in the context of fabric defect detection, is indeed a widely encountered use case. This task involves handling a diverse spectrum of colors and textile types, encompassing a wide range of fabrics. Given the extensive variability in colors, textures, and defect types, fabric defect detection poses a complex and challenging problem in the field of patterned textures inspection. In this article, we propose a knowledge distillation-based approach tailored specifically for addressing the challenge of unsupervised anomaly detection in textures resembling fabrics. Our method aims to redefine the recently introduced reverse distillation approach, which advocates for an encoder-decoder design to mitigate classifier bias and to prevent the student from reconstructing anomalies. In this study, we present a new reverse distillation technique for the specific task of fabric defect detection. Our approach involves a meticulous design selection that strategically highlights high-level features. To demonstrate the capabilities of our approach both in terms of performance and inference speed, we conducted a series of experiments on multiple texture datasets, including MVTEC AD, AITEX, and TILDA, alongside conducting experiments on a dataset acquired from a textile manufacturing facility. The main contributions of this paper are the following: a robust texture anomaly detector utilizing a reverse knowledge-distillation technique suitable for both anomaly detection and domain generalization and a novel dataset encompassing a diverse range of fabrics and defects.

논문 ID: document/1004952

CEKD: Cross-Modal Edge-Privileged Knowledge Distillation for Semantic Scene Understanding Using Only Thermal Images

Semantic scene understanding using thermal images has received great attention due to the advantage that thermal imaging cameras could see in challenging illumination conditions. However, thermal images are lack of color information and the edges in thermal images are often blurred, making them not very suitable to be directly used by existing semantic segmentation networks that are designed with RGB images. To address this problem, we propose a cross-modal edge-privileged knowledge distillation framework, which utilizes a well-trained RGB-Thermal fusion-based semantic segmentation network with edge-privileged information as the teacher, to guide the training of a semantic segmentation network as the student. The student network only uses thermal images. The experimental results on a public dataset demonstrate that under the guidance of the teacher, the student network achieves superior performance over the state of the arts using only thermal images.

arXiv:2401.00797공식 코드 · Papers with Code

Curriculum-scheduled Knowledge Distillation from Multiple Pre-trained Teachers for Multi-domain Sequential Recommendation

Pre-trained recommendation models (PRMs) have received increasing interest recently. However, their intrinsically heterogeneous model structure, huge model size and computation cost hinder their adoptions in practical recommender systems. Hence, it is highly essential to explore how to use different pre-trained recommendation models efficiently in real-world systems. In this paper, we propose a novel curriculum-scheduled knowledge distillation from multiple pre-trained teachers for multi-domain sequential recommendation, called CKD-MDSR, which takes full advantages of different PRMs as multiple teacher models to boost a small student recommendation model, integrating the knowledge across multiple domains from PRMs. Specifically, CKD-MDSR first adopts curriculum-scheduled user behavior sequence sampling and distills informative knowledge jointly from the representative PRMs such as UniSRec and Recformer. Then, the knowledge from the above PRMs are selectively integrated into the student model in consideration of their confidence and consistency. Finally, we verify the proposed method on multi-domain sequential recommendation and further demonstrate its universality with multiple types of student models, including feature interaction and graph based recommendation models. Extensive experiments on five real-world datasets demonstrate the effectiveness and efficiency of CKD-MDSR, which can be viewed as an efficient shortcut using PRMs in real-world systems.

arXiv:2401.00793공식 코드 · Papers with Code

SecFormer: Fast and Accurate Privacy-Preserving Inference for Transformer Models via SMPC

With the growing use of Transformer models hosted on cloud platforms to offer inference services, privacy concerns are escalating, especially concerning sensitive data like investment plans and bank account details. Secure Multi-Party Computing (SMPC) emerges as a promising solution to protect the privacy of inference data and model parameters. However, the application of SMPC in Privacy-Preserving Inference (PPI) for Transformer models often leads to considerable slowdowns or declines in performance. This is largely due to the multitude of nonlinear operations in the Transformer architecture, which are not well-suited to SMPC and difficult to circumvent or optimize effectively. To address this concern, we introduce a comprehensive PPI framework called SecFormer to achieve fast and accurate PPI for Transformer models. We successfully eliminate the high-cost exponential and maximum operations in PPI without sacrificing model performance and develop a suite of efficient SMPC protocols by employing suitable numerical computation methods to boost other complex nonlinear functions in PPI, including GeLU, LayerNorm, and a redesigned Softmax. Our extensive experiments reveal that SecFormer outperforms MPCFormer in performance, showing improvements of $3.4\%$ and $24.7\%$ for BERT$_{\text{BASE}}$ and BERT$_{\text{LARGE}}$, respectively. In terms of efficiency, SecFormer is 3.57 and 3.58 times faster than PUMA for BERT$_{\text{BASE}}$ and BERT$_{\text{LARGE}}$, demonstrating its effectiveness and speed.

arXiv:2401.01145공식 코드 · Papers with Code

HAAQI-Net: A Non-intrusive Neural Music Audio Quality Assessment Model for Hearing Aids

This paper introduces HAAQI-Net, a non-intrusive deep learning-based music audio quality assessment model for hearing aid users. Unlike traditional methods like the Hearing Aid Audio Quality Index (HAAQI) that require intrusive reference signal comparisons, HAAQI-Net offers a more accessible and computationally efficient alternative. By utilizing a Bidirectional Long Short-Term Memory (BLSTM) architecture with attention mechanisms and features extracted from the pre-trained BEATs model, it can predict HAAQI scores directly from music audio clips and hearing loss patterns. Experimental results demonstrate HAAQI-Net's effectiveness, achieving a Linear Correlation Coefficient (LCC) of 0.9368 , a Spearman's Rank Correlation Coefficient (SRCC) of 0.9486 , and a Mean Squared Error (MSE) of 0.0064 and inference time significantly reduces from 62.52 to 2.54 seconds. To address computational overhead, a knowledge distillation strategy was applied, reducing parameters by 75.85% and inference time by 96.46%, while maintaining strong performance (LCC: 0.9071 , SRCC: 0.9307 , MSE: 0.0091 ). To expand its capabilities, HAAQI-Net was adapted to predict subjective human scores like the Mean Opinion Score (MOS) through fine-tuning. This adaptation significantly improved prediction accuracy, validated through statistical analysis. Furthermore, the robustness of HAAQI-Net was evaluated under varying Sound Pressure Level (SPL) conditions, revealing optimal performance at a reference SPL of 65 dB, with accuracy gradually decreasing as SPL deviated from this point. The advancements in subjective score prediction, SPL robustness, and computational efficiency position HAAQI-Net as a scalable solution for music audio quality assessment in hearing aid applications, contributing to efficient and accurate models in audio signal processing and hearing aid technology.

arXiv:2401.00523

Compressing Deep Image Super-resolution Models

Deep learning techniques have been applied in the context of image super-resolution (SR), achieving remarkable advances in terms of reconstruction performance. Existing techniques typically employ highly complex model structures which result in large model sizes and slow inference speeds. This often leads to high energy consumption and restricts their adoption for practical applications. To address this issue, this work employs a three-stage workflow for compressing deep SR models which significantly reduces their memory requirement. Restoration performance has been maintained through teacher-student knowledge distillation using a newly designed distillation loss. We have applied this approach to two popular image super-resolution networks, SwinIR and EDSR, to demonstrate its effectiveness. The resulting compact models, SwinIRmini and EDSRmini, attain an 89% and 96% reduction in both model size and floating-point operations (FLOPs) respectively, compared to their original versions. They also retain competitive super-resolution performance compared to their original models and other commonly used SR approaches. The source code and pre-trained models for these two lightweight SR approaches are released at https://pikapi22.github.io/CDISM/.