Knowledge distillation is a popular technique to transfer knowledge from large teacher models to a small student model. Typically, the student learns to imitate the teacher by minimizing the KL divergence of its output distribution with the teacher's output distribution. In this work, we argue that such a learning objective is sub-optimal because there exists a discrepancy between the teacher's output distribution and the ground truth label distribution. Therefore, forcing the student to blindly imitate the unreliable teacher output distribution leads to inferior performance. To this end, we propose a novel knowledge distillation objective PTLoss by first representing the vanilla KL-based distillation loss function via a Maclaurin series and then perturbing the leading-order terms in this series. This perturbed loss implicitly transforms the original teacher into a proxy teacher with a distribution closer to the ground truth distribution. We establish the theoretical connection between this "distribution closeness" and the student model generalizability, which enables us to select the PTLoss's perturbation coefficients in a principled way. Extensive experiments on five datasets demonstrate PTLoss can significantly improve the distillation effectiveness for teachers of various scales.
In everyday life, humans often plan their actions by following step-by-step instructions in the form of goal-oriented scripts. Previous work has exploited language models (LMs) to plan for abstract goals of stereotypical activities (e.g., "make a cake"), but leaves more specific goals with multi-facet constraints understudied (e.g., "make a cake for diabetics"). In this paper, we define the task of constrained language planning for the first time. We propose an overgenerate-then-filter approach to improve large language models (LLMs) on this task, and use it to distill a novel constrained language planning dataset, CoScript, which consists of 55,000 scripts. Empirical results demonstrate that our method significantly improves the constrained language planning ability of LLMs, especially on constraint faithfulness. Furthermore, CoScript is demonstrated to be quite effective in endowing smaller LMs with constrained language planning ability.
Federated noisy label learning (FNLL) is emerging as a promising tool for privacy-preserving multi-source decentralized learning. Existing research, relying on the assumption of class-balanced global data, might be incapable to model complicated label noise, especially in medical scenarios. In this paper, we first formulate a new and more realistic federated label noise problem where global data is class-imbalanced and label noise is heterogeneous, and then propose a two-stage framework named FedNoRo for noise-robust federated learning. Specifically, in the first stage of FedNoRo, per-class loss indicators followed by Gaussian Mixture Model are deployed for noisy client identification. In the second stage, knowledge distillation and a distance-aware aggregation function are jointly adopted for noise-robust federated model updating. Experimental results on the widely-used ICH and ISIC2019 datasets demonstrate the superiority of FedNoRo against the state-of-the-art FNLL methods for addressing class imbalance and label noise heterogeneity in real-world FL scenarios.
Text image machine translation (TIMT) has been widely used in various real-world applications, which translates source language texts in images into another target language sentence. Existing methods on TIMT are mainly divided into two categories: the recognition-then-translation pipeline model and the end-to-end model. However, how to transfer knowledge from the pipeline model into the end-to-end model remains an unsolved problem. In this paper, we propose a novel Multi-Teacher Knowledge Distillation (MTKD) method to effectively distillate knowledge into the end-to-end TIMT model from the pipeline model. Specifically, three teachers are utilized to improve the performance of the end-to-end TIMT model. The image encoder in the end-to-end TIMT model is optimized with the knowledge distillation guidance from the recognition teacher encoder, while the sequential encoder and decoder are improved by transferring knowledge from the translation sequential and decoder teacher models. Furthermore, both token and sentence-level knowledge distillations are incorporated to better boost the translation performance. Extensive experimental results show that our proposed MTKD effectively improves the text image translation performance and outperforms existing end-to-end and pipeline models with fewer parameters and less decoding time, illustrating that MTKD can take advantage of both pipeline and end-to-end models.
Existing knowledge distillation works for semantic segmentation mainly focus on transferring high-level contextual knowledge from teacher to student. However, low-level texture knowledge is also of vital importance for characterizing the local structural pattern and global statistical property, such as boundary, smoothness, regularity and color contrast, which may not be well addressed by high-level deep features. In this paper, we are intended to take full advantage of both structural and statistical texture knowledge and propose a novel Structural and Statistical Texture Knowledge Distillation (SSTKD) framework for semantic segmentation. Specifically, for structural texture knowledge, we introduce a Contourlet Decomposition Module (CDM) that decomposes low-level features with iterative Laplacian pyramid and directional filter bank to mine the structural texture knowledge. For statistical knowledge, we propose a Denoised Texture Intensity Equalization Module (DTIEM) to adaptively extract and enhance statistical texture knowledge through heuristics iterative quantization and denoised operation. Finally, each knowledge learning is supervised by an individual loss function, forcing the student network to mimic the teacher better from a broader perspective. Experiments show that the proposed method achieves state-of-the-art performance on Cityscapes, Pascal VOC 2012 and ADE20K datasets.
Human Activity Recognition is an important task in many human-computer collaborative scenarios, whilst having various practical applications. Although uni-modal approaches have been extensively studied, they suffer from data quality and require modality-specific feature engineering, thus not being robust and effective enough for real-world deployment. By utilizing various sensors, Multi-modal Human Activity Recognition could utilize the complementary information to build models that can generalize well. While deep learning methods have shown promising results, their potential in extracting salient multi-modal spatial-temporal features and better fusing complementary information has not been fully explored. Also, reducing the complexity of the multi-modal approach for edge deployment is another problem yet to resolve. To resolve the issues, a knowledge distillation-based Multi-modal Mid-Fusion approach, DMFT, is proposed to conduct informative feature extraction and fusion to resolve the Multi-modal Human Activity Recognition task efficiently. DMFT first encodes the multi-modal input data into a unified representation. Then the DMFT teacher model applies an attentive multi-modal spatial-temporal transformer module that extracts the salient spatial-temporal features. A temporal mid-fusion module is also proposed to further fuse the temporal features. Then the knowledge distillation method is applied to transfer the learned representation from the teacher model to a simpler DMFT student model, which consists of a lite version of the multi-modal spatial-temporal transformer module, to produce the results. Evaluation of DMFT was conducted on two public multi-modal human activity recognition datasets with various state-of-the-art approaches. The experimental results demonstrate that the model achieves competitive performance in terms of effectiveness, scalability, and robustness.
There are various optimization techniques in the realm of 3D, including point cloud-based approaches that use mesh, texture, and voxels which optimize how you store, and how do calculate in 3D. These techniques employ methods such as feed-forward networks, 3D convolutions, graph neural networks, transformers, and sparse tensors. However, the field of 3D is one of the most computationally expensive fields, and these methods have yet to achieve their full potential due to their large capacity, complexity, and computation limits. This paper proposes the application of knowledge distillation techniques, especially for sparse tensors in 3D deep learning, to reduce model sizes while maintaining performance. We analyze and purpose different loss functions, including standard methods and combinations of various losses, to simulate the performance of state-of-the-art models of different Sparse Convolutional NNs. Our experiments are done on the standard ScanNet V2 dataset, and we achieved around 2.6\% mIoU difference with a 4 times smaller model and around 8\% with a 16 times smaller model on the latest state-of-the-art spacio-temporal convents based models.
Knowledge distillation is an effective paradigm for boosting the performance of pocket-size model, especially when multiple teacher models are available, the student would break the upper limit again. However, it is not economical to train diverse teacher models for the disposable distillation. In this paper, we introduce a new concept dubbed Avatars for distillation, which are the inference ensemble models derived from the teacher. Concretely, (1) For each iteration of distillation training, various Avatars are generated by a perturbation transformation. We validate that Avatars own higher upper limit of working capacity and teaching ability, aiding the student model in learning diverse and receptive knowledge perspectives from the teacher model. (2) During the distillation, we propose an uncertainty-aware factor from the variance of statistical differences between the vanilla teacher and Avatars, to adjust Avatars' contribution on knowledge transfer adaptively. Avatar Knowledge Distillation AKD is fundamentally different from existing methods and refines with the innovative view of unequal training. Comprehensive experiments demonstrate the effectiveness of our Avatars mechanism, which polishes up the state-of-the-art distillation methods for dense prediction without more extra computational cost. The AKD brings at most 0.7 AP gains on COCO 2017 for Object Detection and 1.83 mIoU gains on Cityscapes for Semantic Segmentation, respectively. Code is available at https://github.com/Gumpest/AvatarKD.
Large language models (LMs) beyond a certain scale, demonstrate the emergent capability of generating free-text rationales for their predictions via chain-of-thought (CoT) prompting. While CoT can yield dramatically improved performance, such gains are only observed for sufficiently large LMs. Even more concerning, there is little guarantee that the generated rationales are consistent with LM's predictions or faithfully justify the decisions. In this work, we propose a faithful knowledge distillation method to learn a small, self-consistent CoT model from a teacher model that is orders of magnitude larger. To form better supervision, we elicit rationales supporting the gold answers from a large LM (teacher) by contrastive decoding, which encourages the teacher to generate tokens that become more plausible only when the answer is considered. To ensure faithful distillation, we use the teacher-generated rationales to learn a student LM with a counterfactual reasoning objective, which prevents the student from ignoring the rationales to make inconsistent predictions. Experiments show that, while yielding comparable end-task performance, our method can generate CoT rationales that are more faithful than baselines do. Further analysis suggests that such a model respects the rationales more when making decisions; thus, we can improve its performance more by refining its rationales.
Deep learning and remote sensing techniques have significantly advanced water monitoring abilities; however, the need for annotated data remains a challenge. This is particularly problematic in wetland detection, where water extent varies over time and space, demanding multiple annotations for the same area. In this paper, we present DeepAqua, a self-supervised deep learning model that leverages knowledge distillation (a.k.a. teacher-student model) to eliminate the need for manual annotations during the training phase. We utilize the Normalized Difference Water Index (NDWI) as a teacher model to train a Convolutional Neural Network (CNN) for segmenting water from Synthetic Aperture Radar (SAR) images, and to train the student model, we exploit cases where optical- and radar-based water masks coincide, enabling the detection of both open and vegetated water surfaces. DeepAqua represents a significant advancement in computer vision techniques by effectively training semantic segmentation models without any manually annotated data. Experimental results show that DeepAqua outperforms other unsupervised methods by improving accuracy by 7%, Intersection Over Union by 27%, and F1 score by 14%. This approach offers a practical solution for monitoring wetland water extent changes without needing ground truth data, making it highly adaptable and scalable for wetland conservation efforts.
Modern Natural Language Generation (NLG) models come with massive computational and storage requirements. In this work, we study the potential of compressing them, which is crucial for real-world applications serving millions of users. We focus on Knowledge Distillation (KD) techniques, in which a small student model learns to imitate a large teacher model, allowing to transfer knowledge from the teacher to the student. In contrast to much of the previous work, our goal is to optimize the model for a specific NLG task and a specific dataset. Typically in real-world applications, in addition to labeled data there is abundant unlabeled task-specific data, which is crucial for attaining high compression rates via KD. In this work, we conduct a systematic study of task-specific KD techniques for various NLG tasks under realistic assumptions. We discuss the special characteristics of NLG distillation and particularly the exposure bias problem. Following, we derive a family of Pseudo-Target (PT) augmentation methods, substantially extending prior work on sequence-level KD. We propose the Joint-Teaching method, which applies word-level KD to multiple PTs generated by both the teacher and the student. Finally, we validate our findings in an extreme setup with no labeled examples using GPT-4 as the teacher. Our study provides practical model design observations and demonstrates the effectiveness of PT training for task-specific KD in NLG.
Recommender systems now consume large-scale data and play a significant role in improving user experience. Graph Neural Networks (GNNs) have emerged as one of the most effective recommender system models because they model the rich relational information. The ever-growing volume of data can make training GNNs prohibitively expensive. To address this, previous attempts propose to train the GNN models incrementally as new data blocks arrive. Feature and structure knowledge distillation techniques have been explored to allow the GNN model to train in a fast incremental fashion while alleviating the catastrophic forgetting problem. However, preserving the same amount of the historical information for all users is sub-optimal since it fails to take into account the dynamics of each user's change of preferences. For the users whose interests shift substantially, retaining too much of the old knowledge can overly constrain the model, preventing it from quickly adapting to the users' novel interests. In contrast, for users who have static preferences, model performance can benefit greatly from preserving as much of the user's long-term preferences as possible. In this work, we propose a novel training strategy that adaptively learns personalized imitation weights for each user to balance the contribution from the recent data and the amount of knowledge to be distilled from previous time periods. We demonstrate the effectiveness of learning imitation weights via a comparison on five diverse datasets for three state-of-art structure distillation based recommender systems. The performance shows consistent improvement over competitive incremental learning techninques.
Fine-tuning large models is highly effective, however, inference can be expensive and produces carbon emissions. Knowledge distillation has been shown to be a practical solution to reduce inference costs, but the distillation process itself requires significant computational resources. Rather than buying or renting GPUs to fine-tune, then distill a large model, an NLP practitioner might instead choose to allocate the available budget to hire annotators and manually label additional fine-tuning data. In this paper, we investigate how to most efficiently use a fixed budget to build a compact model. Through extensive experiments on six diverse tasks, we show that distilling from T5-XXL (11B) to T5-Small (60M) is almost always a cost-efficient strategy compared to annotating more data to directly train a compact model (T5-Small). We further investigate how the optimal budget allocated towards computation varies across scenarios. We will make our code, datasets, annotation cost estimates, and baseline models available as a benchmark to support further work on cost-efficient training of compact models.
Knowledge distillation conducts an effective model compression method while holding some limitations:(1) the feature based distillation methods only focus on distilling the feature map but are lack of transferring the relation of data examples; (2) the relational distillation methods are either limited to the handcrafted functions for relation extraction, such as L2 norm, or weak in inter- and intra- class relation modeling. Besides, the feature divergence of heterogeneous teacher-student architectures may lead to inaccurate relational knowledge transferring. In this work, we propose a novel training framework named Class-Oriented Relational Self Distillation (CORSD) to address the limitations. The trainable relation networks are designed to extract relation of structured data input, and they enable the whole model to better classify samples by transferring the relational knowledge from the deepest layer of the model to shallow layers. Besides, auxiliary classifiers are proposed to make relation networks capture class-oriented relation that benefits classification task. Experiments demonstrate that CORSD achieves remarkable improvements. Compared to baseline, 3.8%, 1.5% and 4.5% averaged accuracy boost can be observed on CIFAR100, ImageNet and CUB-200-2011, respectively.
Detecting players from sports broadcast videos is essential for intelligent event analysis. However, existing methods assume fixed player categories, incapably accommodating the real-world scenarios where categories continue to evolve. Directly fine-tuning these methods on newly emerging categories also exist the catastrophic forgetting due to the non-stationary distribution. Inspired by recent research on incremental object detection (IOD), we propose a Refined Response Distillation (R^2D) method to effectively mitigate catastrophic forgetting for IOD tasks of the players. Firstly, we design a progressive coarse-to-fine distillation region dividing scheme, separating high-value and low-value regions from classification and regression responses for precise and fine-grained regional knowledge distillation. Subsequently, a tailored refined distillation strategy is developed on regions with varying significance to address the performance limitations posed by pronounced feature homogeneity in the IOD tasks of the players. Furthermore, we present the NBA-IOD and Volleyball-IOD datasets as the benchmark and investigate the IOD tasks of the players systematically. Extensive experiments conducted on benchmarks demonstrate that our method achieves state-of-the-art results.The code and datasets are available at https://github.com/beiyan1911/Players-IOD.
Sequential recommendation aims to capture users' dynamic interest and predicts the next item of users' preference. Most sequential recommendation methods use a deep neural network as sequence encoder to generate user and item representations. Existing works mainly center upon designing a stronger sequence encoder. However, few attempts have been made with training an ensemble of networks as sequence encoders, which is more powerful than a single network because an ensemble of parallel networks can yield diverse prediction results and hence better accuracy. In this paper, we present Ensemble Modeling with contrastive Knowledge Distillation for sequential recommendation (EMKD). Our framework adopts multiple parallel networks as an ensemble of sequence encoders and recommends items based on the output distributions of all these networks. To facilitate knowledge transfer between parallel networks, we propose a novel contrastive knowledge distillation approach, which performs knowledge transfer from the representation level via Intra-network Contrastive Learning (ICL) and Cross-network Contrastive Learning (CCL), as well as Knowledge Distillation (KD) from the logits level via minimizing the Kullback-Leibler divergence between the output distributions of the teacher network and the student network. To leverage contextual information, we train the primary masked item prediction task alongside the auxiliary attribute prediction task as a multi-task learning scheme. Extensive experiments on public benchmark datasets show that EMKD achieves a significant improvement compared with the state-of-the-art methods. Besides, we demonstrate that our ensemble method is a generalized approach that can also improve the performance of other sequential recommenders. Our code is available at this link: https://github.com/hw-du/EMKD.
3D point cloud semantic segmentation is one of the fundamental tasks for environmental understanding. Although significant progress has been made in recent years, the performance of classes with few examples or few points is still far from satisfactory. In this paper, we propose a novel multi-to-single knowledge distillation framework for the 3D point cloud semantic segmentation task to boost the performance of those hard classes. Instead of fusing all the points of multi-scans directly, only the instances that belong to the previously defined hard classes are fused. To effectively and sufficiently distill valuable knowledge from multi-scans, we leverage a multilevel distillation framework, i.e., feature representation distillation, logit distillation, and affinity distillation. We further develop a novel instance-aware affinity distillation algorithm for capturing high-level structural knowledge to enhance the distillation efficacy for hard classes. Finally, we conduct experiments on the SemanticKITTI dataset, and the results on both the validation and test sets demonstrate that our method yields substantial improvements compared with the baseline method. The code is available at \Url{https://github.com/skyshoumeng/M2SKD}.
Interactions between humans are diverse and context-dependent, but previous works have treated them as categorical, disregarding the heavy tail of possible interactions. We propose a new paradigm of learning human-human interactions as free text from a single still image, allowing for flexibility in modeling the unlimited space of situations and relationships between people. To overcome the absence of data labelled specifically for this task, we use knowledge distillation applied to synthetic caption data produced by a large language model without explicit supervision. We show that the pseudo-labels produced by this procedure can be used to train a captioning model to effectively understand human-human interactions in images, as measured by a variety of metrics that measure textual and semantic faithfulness and factual groundedness of our predictions. We further show that our approach outperforms SOTA image captioning and situation recognition models on this task. We will release our code and pseudo-labels along with Waldo and Wenda, a manually-curated test set for still image human-human interaction understanding.
Knowledge-Intensive Visual Question Answering (KI-VQA) refers to answering a question about an image whose answer does not lie in the image. This paper presents a new pipeline for KI-VQA tasks, consisting of a retriever and a reader. First, we introduce DEDR, a symmetric dual encoding dense retrieval framework in which documents and queries are encoded into a shared embedding space using uni-modal (textual) and multi-modal encoders. We introduce an iterative knowledge distillation approach that bridges the gap between the representation spaces in these two encoders. Extensive evaluation on two well-established KI-VQA datasets, i.e., OK-VQA and FVQA, suggests that DEDR outperforms state-of-the-art baselines by 11.6% and 30.9% on OK-VQA and FVQA, respectively. Utilizing the passages retrieved by DEDR, we further introduce MM-FiD, an encoder-decoder multi-modal fusion-in-decoder model, for generating a textual answer for KI-VQA tasks. MM-FiD encodes the question, the image, and each retrieved passage separately and uses all passages jointly in its decoder. Compared to competitive baselines in the literature, this approach leads to 5.5% and 8.5% improvements in terms of question answering accuracy on OK-VQA and FVQA, respectively.
Estimating the layout of a room from a single-shot panoramic image is important in virtual/augmented reality and furniture layout simulation. This involves identifying three-dimensional (3D) geometry, such as the location of corners and boundaries, and performing 3D reconstruction. However, occlusion is a common issue that can negatively impact room layout estimation, and this has not been thoroughly studied to date. It is possible to obtain 3D shape information of rooms as drawings of buildings and coordinates of corners from image datasets, thus we propose providing both 2D panoramic and 3D information to a model to effectively deal with occlusion. However, simply feeding 3D information to a model is not sufficient to utilize the shape information for an occluded area. Therefore, we improve the model by introducing 3D Intersection over Union (IoU) loss to effectively use 3D information. In some cases, drawings are not available or the construction deviates from a drawing. Considering such practical cases, we propose a method for distilling knowledge from a model trained with both images and 3D information to a model that takes only images as input. The proposed model, which is called Shape-Net, achieves state-of-the-art (SOTA) performance on benchmark datasets. We also confirmed its effectiveness in dealing with occlusion through significantly improved accuracy on images with occlusion compared with existing models.