PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2310.07585공식 코드 · Papers with Code

A Discrepancy Aware Framework for Robust Anomaly Detection

Defect detection is a critical research area in artificial intelligence. Recently, synthetic data-based self-supervised learning has shown great potential on this task. Although many sophisticated synthesizing strategies exist, little research has been done to investigate the robustness of models when faced with different strategies. In this paper, we focus on this issue and find that existing methods are highly sensitive to them. To alleviate this issue, we present a Discrepancy Aware Framework (DAF), which demonstrates robust performance consistently with simple and cheap strategies across different anomaly detection benchmarks. We hypothesize that the high sensitivity to synthetic data of existing self-supervised methods arises from their heavy reliance on the visual appearance of synthetic data during decoding. In contrast, our method leverages an appearance-agnostic cue to guide the decoder in identifying defects, thereby alleviating its reliance on synthetic appearance. To this end, inspired by existing knowledge distillation methods, we employ a teacher-student network, which is trained based on synthesized outliers, to compute the discrepancy map as the cue. Extensive experiments on two challenging datasets prove the robustness of our method. Under the simple synthesis strategies, it outperforms existing methods by a large margin. Furthermore, it also achieves the state-of-the-art localization performance. Code is available at: https://github.com/caiyuxuan1120/DAF.

arXiv:2310.07554공식 코드 · Papers with Code

Retrieve Anything To Augment Large Language Models

Large language models (LLMs) face significant challenges stemming from their inherent limitations in knowledge, memory, alignment, and action. These challenges cannot be addressed by LLMs alone, but should rely on assistance from the external world, such as knowledge base, memory store, demonstration examples, and tools. Retrieval augmentation stands as a vital mechanism for bridging the gap between LLMs and the external assistance. However, conventional methods encounter two pressing issues. On the one hand, the general-purpose retrievers are not properly optimized for the retrieval augmentation of LLMs. On the other hand, the task-specific retrievers lack the required versatility, hindering their performance across the diverse retrieval augmentation scenarios. In this work, we present a novel approach, the LLM-Embedder, which comprehensively supports the diverse retrieval augmentation needs of LLMs with one unified embedding model. Training such a unified model is non-trivial, as various retrieval tasks aim to capture distinct semantic relationships, often subject to mutual interference. To address this challenge, we systematically optimize our training methodology. This includes reward formulation based on LLMs' feedback, the stabilization of knowledge distillation, multi-task fine-tuning with explicit instructions, and homogeneous in-batch negative sampling. These optimization strategies contribute to the outstanding empirical performance of the LLM-Embedder. Notably, it yields remarkable enhancements in retrieval augmentation for LLMs, surpassing both general-purpose and task-specific retrievers in various evaluation scenarios. Our checkpoint and source code are publicly available at https://github.com/FlagOpen/FlagEmbedding.

arXiv:2310.07177공식 코드 · Papers with Code

Online Speculative Decoding

Speculative decoding is a pivotal technique to accelerate the inference of large language models (LLMs) by employing a smaller draft model to predict the target model's outputs. However, its efficacy can be limited due to the low predictive accuracy of the draft model, particularly when faced with diverse text inputs and a significant capability gap between the draft and target models. We introduce online speculative decoding to address this challenge. The main idea is to continuously update the (multiple) draft model(s) on observed user query data. Adapting to query distribution mitigates the shifts between the training distribution of the draft model and the query distribution, enabling the draft model to more accurately predict the target model's outputs. We develop a prototype of online speculative decoding based on knowledge distillation and evaluate it using both synthetic and real query data. The results show a substantial increase in the token acceptance rate by 0.1 to 0.65, bringing 1.42x to 2.17x latency reduction. Our code is available at https://github.com/LiuXiaoxuanPKU/OSD.

arXiv:2310.06906공식 코드 · Papers with Code

Distillation Improves Visual Place Recognition for Low Quality Images

Real-time visual localization often utilizes online computing, for which query images or videos are transmitted to remote servers for visual place recognition (VPR). However, limited network bandwidth necessitates image-quality reduction and thus the degradation of global image descriptors, reducing VPR accuracy. We address this issue at the descriptor extraction level with a knowledge-distillation methodology that learns feature representations from high-quality images to extract more discriminative descriptors from low-quality images. Our approach includes the Inter-channel Correlation Knowledge Distillation (ICKD) loss, Mean Squared Error (MSE) loss, and Triplet loss. We validate the proposed losses on multiple VPR methods and datasets subjected to JPEG compression, resolution reduction, and video quantization. We obtain significant improvements in VPR recall rates under all three tested modalities of lowered image quality. Furthermore, we fill a gap in VPR literature on video-based data and its influence on VPR performance. This work contributes to more reliable place recognition in resource-constrained environments.

arXiv:2310.06047

Knowledge Distillation for Anomaly Detection

Unsupervised deep learning techniques are widely used to identify anomalous behaviour. The performance of such methods is a product of the amount of training data and the model size. However, the size is often a limiting factor for the deployment on resource-constrained devices. We present a novel procedure based on knowledge distillation for compressing an unsupervised anomaly detection model into a supervised deployable one and we suggest a set of techniques to improve the detection sensitivity. Compressed models perform comparably to their larger counterparts while significantly reducing the size and memory footprint.

arXiv:2310.06372공식 코드 · Papers with Code

Leveraging Diffusion-Based Image Variations for Robust Training on Poisoned Data

Backdoor attacks pose a serious security threat for training neural networks as they surreptitiously introduce hidden functionalities into a model. Such backdoors remain silent during inference on clean inputs, evading detection due to inconspicuous behavior. However, once a specific trigger pattern appears in the input data, the backdoor activates, causing the model to execute its concealed function. Detecting such poisoned samples within vast datasets is virtually impossible through manual inspection. To address this challenge, we propose a novel approach that enables model training on potentially poisoned datasets by utilizing the power of recent diffusion models. Specifically, we create synthetic variations of all training samples, leveraging the inherent resilience of diffusion models to potential trigger patterns in the data. By combining this generative approach with knowledge distillation, we produce student models that maintain their general performance on the task while exhibiting robust resistance to backdoor triggers.

arXiv:2310.05337

What do larger image classifiers memorise?

The success of modern neural networks has prompted study of the connection between memorisation and generalisation: overparameterised models generalise well, despite being able to perfectly fit (memorise) completely random labels. To carefully study this issue, Feldman proposed a metric to quantify the degree of memorisation of individual training examples, and empirically computed the corresponding memorisation profile of a ResNet on image classification bench-marks. While an exciting first glimpse into what real-world models memorise, this leaves open a fundamental question: do larger neural models memorise more? We present a comprehensive empirical analysis of this question on image classification benchmarks. We find that training examples exhibit an unexpectedly diverse set of memorisation trajectories across model sizes: most samples experience decreased memorisation under larger models, while the rest exhibit cap-shaped or increasing memorisation. We show that various proxies for the Feldman memorization score fail to capture these fundamental trends. Lastly, we find that knowledge distillation, an effective and popular model compression technique, tends to inhibit memorisation, while also improving generalisation. Specifically, memorisation is mostly inhibited on examples with increasing memorisation trajectories, thus pointing at how distillation improves generalisation.

arXiv:2310.04352

Fair Feature Importance Scores for Interpreting Tree-Based Methods and Surrogates

Across various sectors such as healthcare, criminal justice, national security, finance, and technology, large-scale machine learning (ML) and artificial intelligence (AI) systems are being deployed to make critical data-driven decisions. Many have asked if we can and should trust these ML systems to be making these decisions. Two critical components are prerequisites for trust in ML systems: interpretability, or the ability to understand why the ML system makes the decisions it does, and fairness, which ensures that ML systems do not exhibit bias against certain individuals or groups. Both interpretability and fairness are important and have separately received abundant attention in the ML literature, but so far, there have been very few methods developed to directly interpret models with regard to their fairness. In this paper, we focus on arguably the most popular type of ML interpretation: feature importance scores. Inspired by the use of decision trees in knowledge distillation, we propose to leverage trees as interpretable surrogates for complex black-box ML models. Specifically, we develop a novel fair feature importance score for trees that can be used to interpret how each feature contributes to fairness or bias in trees, tree-based ensembles, or tree-based surrogates of any complex ML system. Like the popular mean decrease in impurity for trees, our Fair Feature Importance Score is defined based on the mean decrease (or increase) in group bias. Through simulations as well as real examples on benchmark fairness datasets, we demonstrate that our Fair Feature Importance Score offers valid interpretations for both tree-based ensembles and tree-based surrogates of other ML systems.

arXiv:2310.03669공식 코드 · Papers with Code

LumiNet: The Bright Side of Perceptual Knowledge Distillation

In knowledge distillation literature, feature-based methods have dominated due to their ability to effectively tap into extensive teacher models. In contrast, logit-based approaches, which aim to distill `dark knowledge' from teachers, typically exhibit inferior performance compared to feature-based methods. To bridge this gap, we present LumiNet, a novel knowledge distillation algorithm designed to enhance logit-based distillation. We introduce the concept of 'perception', aiming to calibrate logits based on the model's representation capability. This concept addresses overconfidence issues in logit-based distillation method while also introducing a novel method to distill knowledge from the teacher. It reconstructs the logits of a sample/instances by considering relationships with other samples in the batch. LumiNet excels on benchmarks like CIFAR-100, ImageNet, and MSCOCO, outperforming leading feature-based methods, e.g., compared to KD with ResNet18 and MobileNetV2 on ImageNet, it shows improvements of 1.5% and 2.05%, respectively.

arXiv:2310.03188

Talking Models: Distill Pre-trained Knowledge to Downstream Models via Interactive Communication

Many recent breakthroughs in machine learning have been enabled by the pre-trained foundation models. By scaling up model parameters, training data, and computation resources, foundation models have significantly advanced the state-of-the-art in many applications. However, it is still an open question of how to use these models to perform downstream tasks efficiently. Knowledge distillation (KD) has been explored to tackle this challenge. KD transfers knowledge from a large teacher model to a smaller student model. While KD has been successful in improving student model performance, recent research has discovered that a powerful teacher does not necessarily lead to a powerful student, due to their huge capacity gap. In addition, the potential distribution shifts between the pre-training data and downstream tasks can make knowledge transfer in KD sub-optimal for improving downstream task performance. In this paper, we extend KD with an interactive communication process to help students of downstream tasks learn effectively from pre-trained foundation models. Our design is inspired by the way humans learn from teachers who can explain knowledge in a way that meets the students' needs. Specifically, we let each model (i.e., student and teacher) train two components: (1) an encoder encoding the model's hidden states to a message and (2) a decoder decoding any messages to its own hidden states. With encoder and decoder, not only can the teacher transfer rich information by encoding its hidden states, but also the student can send messages with information of downstream tasks to the teacher. Therefore, knowledge passing from teacher to student can be tailored to the student's capacity and downstream tasks' distributions. We conducted experiments on benchmark datasets to show that our communication mechanism outperforms state-of-the-art distillation techniques.

arXiv:2310.02594

I$^2$KD-SLU: An Intra-Inter Knowledge Distillation Framework for Zero-Shot Cross-Lingual Spoken Language Understanding

Spoken language understanding (SLU) typically includes two subtasks: intent detection and slot filling. Currently, it has achieved great success in high-resource languages, but it still remains challenging in low-resource languages due to the scarcity of labeled training data. Hence, there is a growing interest in zero-shot cross-lingual SLU. Despite of the success of existing zero-shot cross-lingual SLU models, most of them neglect to achieve the mutual guidance between intent and slots. To address this issue, we propose an Intra-Inter Knowledge Distillation framework for zero-shot cross-lingual Spoken Language Understanding (I$^2$KD-SLU) to model the mutual guidance. Specifically, we not only apply intra-knowledge distillation between intent predictions or slot predictions of the same utterance in different languages, but also apply inter-knowledge distillation between intent predictions and slot predictions of the same utterance. Our experimental results demonstrate that our proposed framework significantly improves the performance compared with the strong baselines and achieves the new state-of-the-art performance on the MultiATIS++ dataset, obtaining a significant improvement over the previous best model in overall accuracy.

arXiv:2310.02572

Improving Knowledge Distillation with Teacher's Explanation

Knowledge distillation (KD) improves the performance of a low-complexity student model with the help of a more powerful teacher. The teacher in KD is a black-box model, imparting knowledge to the student only through its predictions. This limits the amount of transferred knowledge. In this work, we introduce a novel Knowledge Explaining Distillation (KED) framework, which allows the student to learn not only from the teacher's predictions but also from the teacher's explanations. We propose a class of superfeature-explaining teachers that provide explanation over groups of features, along with the corresponding student model. We also present a method for constructing the superfeatures. We then extend KED to reduce complexity in convolutional neural networks, to allow augmentation with hidden-representation distillation methods, and to work with a limited amount of training data using chimeric sets. Our experiments over a variety of datasets show that KED students can substantially outperform KD students of similar complexity.

arXiv:2310.02549

Heterogeneous Federated Learning Using Knowledge Codistillation

Federated Averaging, and many federated learning algorithm variants which build upon it, have a limitation: all clients must share the same model architecture. This results in unused modeling capacity on many clients, which limits model performance. To address this issue, we propose a method that involves training a small model on the entire pool and a larger model on a subset of clients with higher capacity. The models exchange information bidirectionally via knowledge distillation, utilizing an unlabeled dataset on a server without sharing parameters. We present two variants of our method, which improve upon federated averaging on image classification and language modeling tasks. We show this technique can be useful even if only out-of-domain or limited in-domain distillation data is available. Additionally, the bi-directional knowledge distillation allows for domain transfer between the models when different pool populations introduce domain shift.

arXiv:2310.02421

Can a student Large Language Model perform as well as it's teacher?

The burgeoning complexity of contemporary deep learning models, while achieving unparalleled accuracy, has inadvertently introduced deployment challenges in resource-constrained environments. Knowledge distillation, a technique aiming to transfer knowledge from a high-capacity "teacher" model to a streamlined "student" model, emerges as a promising solution to this dilemma. This paper provides a comprehensive overview of the knowledge distillation paradigm, emphasizing its foundational principles such as the utility of soft labels and the significance of temperature scaling. Through meticulous examination, we elucidate the critical determinants of successful distillation, including the architecture of the student model, the caliber of the teacher, and the delicate balance of hyperparameters. While acknowledging its profound advantages, we also delve into the complexities and challenges inherent in the process. Our exploration underscores knowledge distillation's potential as a pivotal technique in optimizing the trade-off between model performance and deployment efficiency.

arXiv:2310.01777공식 코드 · Papers with Code

SEA: Sparse Linear Attention with Estimated Attention Mask

The transformer architecture has driven breakthroughs in recent years on tasks which require modeling pairwise relationships between sequential elements, as is the case in natural language understanding. However, long seqeuences pose a problem due to the quadratic complexity of the attention operation. Previous research has aimed to lower the complexity by sparsifying or linearly approximating the attention matrix. Yet, these approaches cannot straightforwardly distill knowledge from a teacher's attention matrix and often require complete retraining from scratch. Furthermore, previous sparse and linear approaches lose interpretability if they cannot produce full attention matrices. To address these challenges, we propose SEA: Sparse linear attention with an Estimated Attention mask. SEA estimates the attention matrix with linear complexity via kernel-based linear attention, then subsequently creates a sparse attention matrix with a top-k selection to perform a sparse attention operation. For language modeling tasks (Wikitext2), previous linear and sparse attention methods show roughly two-fold worse perplexity scores over the quadratic OPT-1.3B baseline, while SEA achieves better perplexity than OPT-1.3B, using roughly half the memory of OPT-1.3B, providing interpretable attention matrix. We believe that our work will have a large practical impact, as it opens the possibility of running large transformers on resource-limited devices with less memory.

arXiv:2310.01035

Learnable Cross-modal Knowledge Distillation for Multi-modal Learning with Missing Modality

The problem of missing modalities is both critical and non-trivial to be handled in multi-modal models. It is common for multi-modal tasks that certain modalities contribute more compared to other modalities, and if those important modalities are missing, the model performance drops significantly. Such fact remains unexplored by current multi-modal approaches that recover the representation from missing modalities by feature reconstruction or blind feature aggregation from other modalities, instead of extracting useful information from the best performing modalities. In this paper, we propose a Learnable Cross-modal Knowledge Distillation (LCKD) model to adaptively identify important modalities and distil knowledge from them to help other modalities from the cross-modal perspective for solving the missing modality issue. Our approach introduces a teacher election procedure to select the most ``qualified'' teachers based on their single modality performance on certain tasks. Then, cross-modal knowledge distillation is performed between teacher and student modalities for each task to push the model parameters to a point that is beneficial for all tasks. Hence, even if the teacher modalities for certain tasks are missing during testing, the available student modalities can accomplish the task well enough based on the learned knowledge from their automatically elected teacher modalities. Experiments on the Brain Tumour Segmentation Dataset 2018 (BraTS2018) shows that LCKD outperforms other methods by a considerable margin, improving the state-of-the-art performance by 3.61% for enhancing tumour, 5.99% for tumour core, and 3.76% for whole tumour in terms of segmentation Dice score.

arXiv:2310.00369

Distilling Inductive Bias: Knowledge Distillation Beyond Model Compression

With the rapid development of computer vision, Vision Transformers (ViTs) offer the tantalizing prospect of unified information processing across visual and textual domains. But due to the lack of inherent inductive biases in ViTs, they require enormous amount of data for training. To make their applications practical, we introduce an innovative ensemble-based distillation approach distilling inductive bias from complementary lightweight teacher models. Prior systems relied solely on convolution-based teaching. However, this method incorporates an ensemble of light teachers with different architectural tendencies, such as convolution and involution, to instruct the student transformer jointly. Because of these unique inductive biases, instructors can accumulate a wide range of knowledge, even from readily identifiable stored datasets, which leads to enhanced student performance. Our proposed framework also involves precomputing and storing logits in advance, essentially the unnormalized predictions of the model. This optimization can accelerate the distillation process by eliminating the need for repeated forward passes during knowledge distillation, significantly reducing the computational burden and enhancing efficiency.

arXiv:2310.00258공식 코드 · Papers with Code

NAYER: Noisy Layer Data Generation for Efficient and Effective Data-free Knowledge Distillation

Data-Free Knowledge Distillation (DFKD) has made significant recent strides by transferring knowledge from a teacher neural network to a student neural network without accessing the original data. Nonetheless, existing approaches encounter a significant challenge when attempting to generate samples from random noise inputs, which inherently lack meaningful information. Consequently, these models struggle to effectively map this noise to the ground-truth sample distribution, resulting in prolonging training times and low-quality outputs. In this paper, we propose a novel Noisy Layer Generation method (NAYER) which relocates the random source from the input to a noisy layer and utilizes the meaningful constant label-text embedding (LTE) as the input. LTE is generated by using the language model once, and then it is stored in memory for all subsequent training processes. The significance of LTE lies in its ability to contain substantial meaningful inter-class information, enabling the generation of high-quality samples with only a few training steps. Simultaneously, the noisy layer plays a key role in addressing the issue of diversity in sample generation by preventing the model from overemphasizing the constrained label information. By reinitializing the noisy layer in each iteration, we aim to facilitate the generation of diverse samples while still retaining the method's efficiency, thanks to the ease of learning provided by LTE. Experiments carried out on multiple datasets demonstrate that our NAYER not only outperforms the state-of-the-art methods but also achieves speeds 5 to 15 times faster than previous approaches. The code is available at https://github.com/tmtuan1307/nayer.

arXiv:2310.00096

Towards Few-Call Model Stealing via Active Self-Paced Knowledge Distillation and Diffusion-Based Image Generation

Diffusion models showcased strong capabilities in image synthesis, being used in many computer vision tasks with great success. To this end, we propose to explore a new use case, namely to copy black-box classification models without having access to the original training data, the architecture, and the weights of the model, \ie~the model is only exposed through an inference API. More specifically, we can only observe the (soft or hard) labels for some image samples passed as input to the model. Furthermore, we consider an additional constraint limiting the number of model calls, mostly focusing our research on few-call model stealing. In order to solve the model extraction task given the applied restrictions, we propose the following framework. As training data, we create a synthetic data set (called proxy data set) by leveraging the ability of diffusion models to generate realistic and diverse images. Given a maximum number of allowed API calls, we pass the respective number of samples through the black-box model to collect labels. Finally, we distill the knowledge of the black-box teacher (attacked model) into a student model (copy of the attacked model), harnessing both labeled and unlabeled data generated by the diffusion model. We employ a novel active self-paced learning framework to make the most of the proxy data during distillation. Our empirical results on two data sets confirm the superiority of our framework over two state-of-the-art methods in the few-call model extraction scenario.

arXiv:2310.00946공식 코드 · Papers with Code

Distilling Influences to Mitigate Prediction Churn in Graph Neural Networks

Models with similar performances exhibit significant disagreement in the predictions of individual samples, referred to as prediction churn. Our work explores this phenomenon in graph neural networks by investigating differences between models differing only in their initializations in their utilized features for predictions. We propose a novel metric called Influence Difference (ID) to quantify the variation in reasons used by nodes across models by comparing their influence distribution. Additionally, we consider the differences between nodes with a stable and an unstable prediction, positing that both equally utilize different reasons and thus provide a meaningful gradient signal to closely match two models even when the predictions for nodes are similar. Based on our analysis, we propose to minimize this ID in Knowledge Distillation, a domain where a new model should closely match an established one. As an efficient approximation, we introduce DropDistillation (DD) that matches the output for a graph perturbed by edge deletions. Our empirical evaluation of six benchmark datasets for node classification validates the differences in utilized features. DD outperforms previous methods regarding prediction stability and overall performance in all considered Knowledge Distillation experiments.