PyTorchKR Papers

최신 논문

총 3,892편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2407.21687

Dynamic Object Queries for Transformer-based Incremental Object Detection

Incremental object detection (IOD) aims to sequentially learn new classes, while maintaining the capability to locate and identify old ones. As the training data arrives with annotations only with new classes, IOD suffers from catastrophic forgetting. Prior methodologies mainly tackle the forgetting issue through knowledge distillation and exemplar replay, ignoring the conflict between limited model capacity and increasing knowledge. In this paper, we explore \textit{dynamic object queries} for incremental object detection built on Transformer architecture. We propose the \textbf{Dy}namic object \textbf{Q}uery-based \textbf{DE}tection \textbf{TR}ansformer (DyQ-DETR), which incrementally expands the model representation ability to achieve stability-plasticity tradeoff. First, a new set of learnable object queries are fed into the decoder to represent new classes. These new object queries are aggregated with those from previous phases to adapt both old and new knowledge well. Second, we propose the isolated bipartite matching for object queries in different phases, based on disentangled self-attention. The interaction among the object queries at different phases is eliminated to reduce inter-class confusion. Thanks to the separate supervision and computation over object queries, we further present the risk-balanced partial calibration for effective exemplar replay. Extensive experiments demonstrate that DyQ-DETR significantly surpasses the state-of-the-art methods, with limited parameter overhead. Code will be made publicly available.

arXiv:2407.21416

VIPeR: Visual Incremental Place Recognition with Adaptive Mining and Continual Learning

Visual place recognition (VPR) is an essential component of many autonomous and augmented/virtual reality systems. It enables the systems to robustly localize themselves in large-scale environments. Existing VPR methods demonstrate attractive performance at the cost of heavy pre-training and limited generalizability. When deployed in unseen environments, these methods exhibit significant performance drops. Targeting this issue, we present VIPeR, a novel approach for visual incremental place recognition with the ability to adapt to new environments while retaining the performance of previous environments. We first introduce an adaptive mining strategy that balances the performance within a single environment and the generalizability across multiple environments. Then, to prevent catastrophic forgetting in lifelong learning, we draw inspiration from human memory systems and design a novel memory bank for our VIPeR. Our memory bank contains a sensory memory, a working memory and a long-term memory, with the first two focusing on the current environment and the last one for all previously visited environments. Additionally, we propose a probabilistic knowledge distillation to explicitly safeguard the previously learned knowledge. We evaluate our proposed VIPeR on three large-scale datasets, namely Oxford Robotcar, Nordland, and TartanAir. For comparison, we first set a baseline performance with naive finetuning. Then, several more recent lifelong learning methods are compared. Our VIPeR achieves better performance in almost all aspects with the biggest improvement of 13.65% in average performance.

arXiv:2407.21252

Lifelong Person Search

Person search is the task to localize a query person in gallery datasets of scene images. Existing methods have been mainly developed to handle a single target dataset only, however diverse datasets are continuously given in practical applications of person search. In such cases, they suffer from the catastrophic knowledge forgetting in the old datasets when trained on new datasets. In this paper, we first introduce a novel problem of lifelong person search (LPS) where the model is incrementally trained on the new datasets while preserving the knowledge learned in the old datasets. We propose an end-to-end LPS framework that facilitates the knowledge distillation to enforce the consistency learning between the old and new models by utilizing the prototype features of the foreground persons as well as the hard background proposals in the old domains. Moreover, we also devise the rehearsal-based instance matching to further improve the discrimination ability in the old domains by using the unlabeled person instances additionally. Experimental results demonstrate that the proposed method achieves significantly superior performance of both the detection and re-identification to preserve the knowledge learned in the old domains compared with the existing methods.

arXiv:2407.21515

Learning Effective Representations for Retrieval Using Self-Distillation with Adaptive Relevance Margins

Representation-based retrieval models, so-called biencoders, estimate the relevance of a document to a query by calculating the similarity of their respective embeddings. Current state-of-the-art biencoders are trained using an expensive training regime involving knowledge distillation from a teacher model and batch-sampling. Instead of relying on a teacher model, we contribute a novel parameter-free loss function for self-supervision that exploits the pre-trained language modeling capabilities of the encoder model as a training signal, eliminating the need for batch sampling by performing implicit hard negative mining. We investigate the capabilities of our proposed approach through extensive ablation studies, demonstrating that self-distillation can match the effectiveness of teacher distillation using only 13.5% of the data, while offering a speedup in training time between 3x and 15x compared to parametrized losses. Code and data is made openly available.

arXiv:2407.20584공식 코드 · Papers with Code

Pruning Large Language Models with Semi-Structural Adaptive Sparse Training

The remarkable success of Large Language Models (LLMs) relies heavily on their substantial scale, which poses significant challenges during model deployment in terms of latency and memory consumption. Recently, numerous studies have attempted to compress LLMs using one-shot pruning methods. However, these methods often suffer from considerable performance degradation on complex language understanding tasks, raising concerns about the feasibility of pruning in LLMs. To address this issue, we propose Adaptive Sparse Trainer (AST), a novel and efficient retraining framework tailored for semi-structured sparse models. AST enables models to learn optimal masks during the weight update process without incurring additional computational overhead. Furthermore, we demonstrate that incorporating knowledge distillation significantly improves retraining efficiency and enhances model performance under fixed computational constraints. Additionally, a supplementary set of well-initialized parameters is integrated to further augment the model's efficacy. AST achieves state-of-the-art performance with minimal training cost. When applied to the LLaMA2-7B model, AST reduces the perplexity and zero-shot accuracy gap between dense and 2:4 semi-structured sparse models to 0.6 and 1.16%, respectively, utilizing less than 0.4% of the pretraining tokens and GPU hours. Our work demonstrates the feasibility of deploying semi-structured sparse LLMs and offers a promising alternative for achieving highly compressed models when combined with existing quantization techniques.

arXiv:2407.19409

LLAVADI: What Matters For Multimodal Large Language Models Distillation

The recent surge in Multimodal Large Language Models (MLLMs) has showcased their remarkable potential for achieving generalized intelligence by integrating visual understanding into Large Language Models.Nevertheless, the sheer model size of MLLMs leads to substantial memory and computational demands that hinder their widespread deployment. In this work, we do not propose a new efficient model structure or train small-scale MLLMs from scratch. Instead, we focus on what matters for training small-scale MLLMs through knowledge distillation, which is the first step from the multimodal distillation perspective. Our extensive studies involve training strategies, model choices, and distillation algorithms in the knowledge distillation process. These results show that joint alignment for both tokens and logit alignment plays critical roles in teacher-student frameworks. In addition, we draw a series of intriguing observations from this study. By evaluating different benchmarks and proper strategy, even a 2.7B small-scale model can perform on par with larger models with 7B or 13B parameters. Our code and models will be publicly available for further research.

arXiv:2407.19271

Sewer Image Super-Resolution with Depth Priors and Its Lightweight Network

The Quick-view (QV) technique serves as a primary method for detecting defects within sewerage systems. However, the effectiveness of QV is impeded by the limited visual range of its hardware, resulting in suboptimal image quality for distant portions of the sewer network. Image super-resolution is an effective way to improve image quality and has been applied in a variety of scenes. However, research on super-resolution for sewer images remains considerably unexplored. In response, this study leverages the inherent depth relationships present within QV images and introduces a novel Depth-guided, Reference-based Super-Resolution framework denoted as DSRNet. It comprises two core components: a depth extraction module and a depth information matching module (DMM). DSRNet utilizes the adjacent frames of the low-resolution image as reference images and helps them recover texture information based on the correlation. By combining these modules, the integration of depth priors significantly enhances both visual quality and performance benchmarks. Besides, in pursuit of computational efficiency and compactness, a super-resolution knowledge distillation model based on an attention mechanism is introduced. This mechanism facilitates the acquisition of feature similarity between a more complex teacher model and a streamlined student model, with the latter being a lightweight version of DSRNet. Experimental results demonstrate that DSRNet significantly improves PSNR and SSIM compared with other methods. This study also conducts experiments on sewer defect semantic segmentation, object detection, and classification on the Pipe dataset and Sewer-ML dataset. Experiments show that the method can improve the performance of low-resolution sewer images in these tasks.

arXiv:2407.19820

ActivityCLIP: Enhancing Group Activity Recognition by Mining Complementary Information from Text to Supplement Image Modality

Previous methods usually only extract the image modality's information to recognize group activity. However, mining image information is approaching saturation, making it difficult to extract richer information. Therefore, extracting complementary information from other modalities to supplement image information has become increasingly important. In fact, action labels provide clear text information to express the action's semantics, which existing methods often overlook. Thus, we propose ActivityCLIP, a plug-and-play method for mining the text information contained in the action labels to supplement the image information for enhancing group activity recognition. ActivityCLIP consists of text and image branches, where the text branch is plugged into the image branch (The off-the-shelf image-based method). The text branch includes Image2Text and relation modeling modules. Specifically, we propose the knowledge transfer module, Image2Text, which adapts image information into text information extracted by CLIP via knowledge distillation. Further, to keep our method convenient, we add fewer trainable parameters based on the relation module of the image branch to model interaction relation in the text branch. To show our method's generality, we replicate three representative methods by ActivityCLIP, which adds only limited trainable parameters, achieving favorable performance improvements for each method. We also conduct extensive ablation studies and compare our method with state-of-the-art methods to demonstrate the effectiveness of ActivityCLIP.

arXiv:2407.19405

Logic Distillation: Learning from Code Function by Function for Planning and Decision-making

Large language models (LLMs) have garnered increasing attention owing to their powerful logical reasoning capabilities. Generally, larger LLMs (L-LLMs) that require paid interfaces exhibit significantly superior performance compared to smaller LLMs (S-LLMs) that can be deployed on a variety of devices. Knowledge distillation (KD) aims to empower S-LLMs with the capabilities of L-LLMs, while S-LLMs merely mimic the outputs of L-LLMs, failing to get the powerful logical reasoning capabilities. Consequently, S-LLMs are helpless when it comes to planning and decision-making tasks that require logical reasoning capabilities. To tackle the identified challenges, we propose a novel framework called Logic Distillation (LD). Initially, LD employs L-LLMs to instantiate complex instructions into discrete functions and illustrates their usage to establish a function base. Subsequently, based on the function base, LD fine-tunes S-LLMs to learn the logic employed by L-LLMs in planning and decision-making. During testing, LD utilizes a retriever to identify the top-$K$ relevant functions based on instructions and current states, which will be selected and invoked by S-LLMs. Ultimately, S-LLMs yield planning and decision-making outcomes, function by function. Relevant experiments demonstrate that with the assistance of LD, S-LLMs can achieve outstanding results in planning and decision-making tasks, comparable to, or even surpassing, those of L-LLMs.

arXiv:2407.20062공식 코드 · Papers with Code

SalNAS: Efficient Saliency-prediction Neural Architecture Search with self-knowledge distillation

Recent advancements in deep convolutional neural networks have significantly improved the performance of saliency prediction. However, the manual configuration of the neural network architectures requires domain knowledge expertise and can still be time-consuming and error-prone. To solve this, we propose a new Neural Architecture Search (NAS) framework for saliency prediction with two contributions. Firstly, a supernet for saliency prediction is built with a weight-sharing network containing all candidate architectures, by integrating a dynamic convolution into the encoder-decoder in the supernet, termed SalNAS. Secondly, despite the fact that SalNAS is highly efficient (20.98 million parameters), it can suffer from the lack of generalization. To solve this, we propose a self-knowledge distillation approach, termed Self-KD, that trains the student SalNAS with the weighted average information between the ground truth and the prediction from the teacher model. The teacher model, while sharing the same architecture, contains the best-performing weights chosen by cross-validation. Self-KD can generalize well without the need to compute the gradient in the teacher model, enabling an efficient training system. By utilizing Self-KD, SalNAS outperforms other state-of-the-art saliency prediction models in most evaluation rubrics across seven benchmark datasets while being a lightweight model. The code will be available at https://github.com/chakkritte/SalNAS

arXiv:2407.19610공식 코드 · Papers with Code

Mixture of Modular Experts: Distilling Knowledge from a Multilingual Teacher into Specialized Modular Language Models

This research combines Knowledge Distillation (KD) and Mixture of Experts (MoE) to develop modular, efficient multilingual language models. Key objectives include evaluating adaptive versus fixed alpha methods in KD and comparing modular MoE architectures for handling multi-domain inputs and preventing catastrophic forgetting. KD compresses large language models (LLMs) into smaller, efficient models, while MoE enhances modularity with specialized tasks. Experiments showed similar performance for both KD methods, with marginal improvements from adaptive alpha. A combined loss approach provided more stable learning. The router, trained to classify input sequences into English, French, German, or Python, achieved 99.95% precision, recall, and F1 score, with Logistic Regression being the most effective classifier. Evaluations of modular MoE architectures revealed that Pre-trained Language Experts (PLE) and Joint Expert Embedding Training (JEET) performed similarly, while the MoE with Common Expert (MoE-CE) setup showed slightly lower performance. Including a common expert in MoE-CE improved its performance. Studies on catastrophic forgetting indicated that sequential training led to significant forgetting, while single-session training with balanced batches and the MoE approach mitigated this issue. The MoE architecture preserved knowledge across multiple languages effectively. The research contributes open-sourced resources including the dataset (https://zenodo.org/doi/10.5281/zenodo.12677631), a balanced dataset creation tool (https://github.com/padas-lab-de/multi-language-dataset-creator), and the research codebase (https://github.com/ModMaamari/mixture-modular-experts).

arXiv:2407.19540공식 코드 · Papers with Code

Overcoming Uncertain Incompleteness for Robust Multimodal Sequential Diagnosis Prediction via Curriculum Data Erasing Guided Knowledge Distillation

In this paper, we present NECHO v2, a novel framework designed to enhance the predictive accuracy of multimodal sequential patient diagnoses under uncertain missing visit sequences, a common challenge in real clinical settings. Firstly, we modify NECHO, designed in a diagnosis code-centric fashion, to handle uncertain modality representation dominance under the imperfect data. Secondly, we develop a systematic knowledge distillation by employing the modified NECHO as both teacher and student. It encompasses a modality-wise contrastive and hierarchical distillation, transformer representation random distillation, along with other distillations to align representations between teacher and student tightly and effectively. We also propose curriculum learning guided random data erasing within sequences during both training and distillation of the teacher to lightly simulate scenario with missing visit information, thereby fostering effective knowledge transfer. As a result, NECHO v2 verifies itself by showing robust superiority in multimodal sequential diagnosis prediction under both balanced and imbalanced incomplete settings on multimodal healthcare data.

arXiv:2407.18472

FedUD: Exploiting Unaligned Data for Cross-Platform Federated Click-Through Rate Prediction

Click-through rate (CTR) prediction plays an important role in online advertising platforms. Most existing methods use data from the advertising platform itself for CTR prediction. As user behaviors also exist on many other platforms, e.g., media platforms, it is beneficial to further exploit such complementary information for better modeling user interest and for improving CTR prediction performance. However, due to privacy concerns, data from different platforms cannot be uploaded to a server for centralized model training. Vertical federated learning (VFL) provides a possible solution which is able to keep the raw data on respective participating parties and learn a collaborative model in a privacy-preserving way. However, traditional VFL methods only utilize aligned data with common keys across parties, which strongly restricts their application scope. In this paper, we propose FedUD, which is able to exploit unaligned data, in addition to aligned data, for more accurate federated CTR prediction. FedUD contains two steps. In the first step, FedUD utilizes aligned data across parties like traditional VFL, but it additionally includes a knowledge distillation module. This module distills useful knowledge from the guest party's high-level representations and guides the learning of a representation transfer network. In the second step, FedUD applies the learned knowledge to enrich the representations of the host party's unaligned data such that both aligned and unaligned data can contribute to federated model training. Experiments on two real-world datasets demonstrate the superior performance of FedUD for federated CTR prediction.

arXiv:2407.18534공식 코드 · Papers with Code

Boosting Cross-Domain Point Classification via Distilling Relational Priors from 2D Transformers

Semantic pattern of an object point cloud is determined by its topological configuration of local geometries. Learning discriminative representations can be challenging due to large shape variations of point sets in local regions and incomplete surface in a global perspective, which can be made even more severe in the context of unsupervised domain adaptation (UDA). In specific, traditional 3D networks mainly focus on local geometric details and ignore the topological structure between local geometries, which greatly limits their cross-domain generalization. Recently, the transformer-based models have achieved impressive performance gain in a range of image-based tasks, benefiting from its strong generalization capability and scalability stemming from capturing long range correlation across local patches. Inspired by such successes of visual transformers, we propose a novel Relational Priors Distillation (RPD) method to extract relational priors from the well-trained transformers on massive images, which can significantly empower cross-domain representations with consistent topological priors of objects. To this end, we establish a parameter-frozen pre-trained transformer module shared between 2D teacher and 3D student models, complemented by an online knowledge distillation strategy for semantically regularizing the 3D student model. Furthermore, we introduce a novel self-supervised task centered on reconstructing masked point cloud patches using corresponding masked multi-view image features, thereby empowering the model with incorporating 3D geometric information. Experiments on the PointDA-10 and the Sim-to-Real datasets verify that the proposed method consistently achieves the state-of-the-art performance of UDA for point cloud classification. The source code of this work is available at https://github.com/zou-longkun/RPD.git.

arXiv:2407.18449

Towards A Generalizable Pathology Foundation Model via Unified Knowledge Distillation

Foundation models pretrained on large-scale datasets are revolutionizing the field of computational pathology (CPath). The generalization ability of foundation models is crucial for the success in various downstream clinical tasks. However, current foundation models have only been evaluated on a limited type and number of tasks, leaving their generalization ability and overall performance unclear. To address this gap, we established a most comprehensive benchmark to evaluate the performance of off-the-shelf foundation models across six distinct clinical task types, encompassing a total of 72 specific tasks, including slide-level classification, survival prediction, ROI-tissue classification, ROI retrieval, visual question answering, and report generation. Our findings reveal that existing foundation models excel at certain task types but struggle to effectively handle the full breadth of clinical tasks. To improve the generalization of pathology foundation models, we propose a unified knowledge distillation framework consisting of both expert and self-knowledge distillation, where the former allows the model to learn from the knowledge of multiple expert models, while the latter leverages self-distillation to enable image representation learning via local-global alignment. Based on this framework, we curated a dataset of 96,000 whole slide images (WSIs) and developed a Generalizable Pathology Foundation Model (GPFM). This advanced model was trained on a substantial dataset comprising 190 million images extracted from approximately 72,000 publicly available slides, encompassing 34 major tissue types. Evaluated on the established benchmark, GPFM achieves an impressive average rank of 1.6, with 42 tasks ranked 1st, while the second-best model, UNI, attains an average rank of 3.7, with only 6 tasks ranked 1st.

arXiv:2407.18288공식 코드 · Papers with Code

Leveraging Foundation Models via Knowledge Distillation in Multi-Object Tracking: Distilling DINOv2 Features to FairMOT

Multiple Object Tracking (MOT) is a computer vision task that has been employed in a variety of sectors. Some common limitations in MOT are varying object appearances, occlusions, or crowded scenes. To address these challenges, machine learning methods have been extensively deployed, leveraging large datasets, sophisticated models, and substantial computational resources. Due to practical limitations, access to the above is not always an option. However, with the recent release of foundation models by prominent AI companies, pretrained models have been trained on vast datasets and resources using state-of-the-art methods. This work tries to leverage one such foundation model, called DINOv2, through using knowledge distillation. The proposed method uses a teacher-student architecture, where DINOv2 is the teacher and the FairMOT backbone HRNetv2 W18 is the student. The results imply that although the proposed method shows improvements in certain scenarios, it does not consistently outperform the original FairMOT model. These findings highlight the potential and limitations of applying foundation models in knowledge

arXiv:2407.18248공식 코드 · Papers with Code

Self-Training with Direct Preference Optimization Improves Chain-of-Thought Reasoning

Effective training of language models (LMs) for mathematical reasoning tasks demands high-quality supervised fine-tuning data. Besides obtaining annotations from human experts, a common alternative is sampling from larger and more powerful LMs. However, this knowledge distillation approach can be costly and unstable, particularly when relying on closed-source, proprietary LMs like GPT-4, whose behaviors are often unpredictable. In this work, we demonstrate that the reasoning abilities of small-scale LMs can be enhanced through self-training, a process where models learn from their own outputs. We also show that the conventional self-training can be further augmented by a preference learning algorithm called Direct Preference Optimization (DPO). By integrating DPO into self-training, we leverage preference data to guide LMs towards more accurate and diverse chain-of-thought reasoning. We evaluate our method across various mathematical reasoning tasks using different base models. Our experiments show that this approach not only improves LMs' reasoning performance but also offers a more cost-effective and scalable solution compared to relying on large proprietary LMs.

arXiv:2407.18041공식 코드 · Papers with Code

How to Train the Teacher Model for Effective Knowledge Distillation

Recently, it was shown that the role of the teacher in knowledge distillation (KD) is to provide the student with an estimate of the true Bayes conditional probability density (BCPD). Notably, the new findings propose that the student's error rate can be upper-bounded by the mean squared error (MSE) between the teacher's output and BCPD. Consequently, to enhance KD efficacy, the teacher should be trained such that its output is close to BCPD in MSE sense. This paper elucidates that training the teacher model with MSE loss equates to minimizing the MSE between its output and BCPD, aligning with its core responsibility of providing the student with a BCPD estimate closely resembling it in MSE terms. In this respect, through a comprehensive set of experiments, we demonstrate that substituting the conventional teacher trained with cross-entropy loss with one trained using MSE loss in state-of-the-art KD methods consistently boosts the student's accuracy, resulting in improvements of up to 2.6\%.

arXiv:2407.17909

Separating Novel Features for Logical Anomaly Detection: A Straightforward yet Effective Approach

Vision-based inspection algorithms have significantly contributed to quality control in industrial settings, particularly in addressing structural defects like dent and contamination which are prevalent in mass production. Extensive research efforts have led to the development of related benchmarks such as MVTec AD (Bergmann et al., 2019). However, in industrial settings, there can be instances of logical defects, where acceptable items are found in unsuitable locations or product pairs do not match as expected. Recent methods tackling logical defects effectively employ knowledge distillation to generate difference maps. Knowledge distillation (KD) is used to learn normal data distribution in unsupervised manner. Despite their effectiveness, these methods often overlook the potential false negatives. Excessive similarity between the teacher network and student network can hinder the generation of a suitable difference map for logical anomaly detection. This technical report provides insights on handling potential false negatives by utilizing a simple constraint in KD-based logical anomaly detection methods. We select EfficientAD as a state-of-the-art baseline and apply a margin-based constraint to its unsupervised learning scheme. Applying this constraint, we can improve the AUROC for MVTec LOCO AD by 1.3 %.

arXiv:2407.17620공식 코드 · Papers with Code

CoMoTo: Unpaired Cross-Modal Lesion Distillation Improves Breast Lesion Detection in Tomosynthesis

Digital Breast Tomosynthesis (DBT) is an advanced breast imaging modality that offers superior lesion detection accuracy compared to conventional mammography, albeit at the trade-off of longer reading time. Accelerating lesion detection from DBT using deep learning is hindered by limited data availability and huge annotation costs. A possible solution to this issue could be to leverage the information provided by a more widely available modality, such as mammography, to enhance DBT lesion detection. In this paper, we present a novel framework, CoMoTo, for improving lesion detection in DBT. Our framework leverages unpaired mammography data to enhance the training of a DBT model, improving practicality by eliminating the need for mammography during inference. Specifically, we propose two novel components, Lesion-specific Knowledge Distillation (LsKD) and Intra-modal Point Alignment (ImPA). LsKD selectively distills lesion features from a mammography teacher model to a DBT student model, disregarding background features. ImPA further enriches LsKD by ensuring the alignment of lesion features within the teacher before distilling knowledge to the student. Our comprehensive evaluation shows that CoMoTo is superior to traditional pretraining and image-level KD, improving performance by 7% Mean Sensitivity under low-data setting. Our code is available at https://github.com/Muhammad-Al-Barbary/CoMoTo .