The ongoing biodiversity crisis, driven by factors such as land-use change and global warming, emphasizes the need for effective ecological monitoring methods. Acoustic monitoring of biodiversity has emerged as an important monitoring tool. Detecting human voices in soundscape monitoring projects is useful both for analysing human disturbance and for privacy filtering. Despite significant strides in deep learning in recent years, the deployment of large neural networks on compact devices poses challenges due to memory and latency constraints. Our approach focuses on leveraging knowledge distillation techniques to design efficient, lightweight student models for speech detection in bioacoustics. In particular, we employed the MobileNetV3-Small-Pi model to create compact yet effective student architectures to compare against the larger EcoVAD teacher model, a well-regarded voice detection architecture in eco-acoustic monitoring. The comparative analysis included examining various configurations of the MobileNetV3-Small-Pi derived student models to identify optimal performance. Additionally, a thorough evaluation of different distillation techniques was conducted to ascertain the most effective method for model selection. Our findings revealed that the distilled models exhibited comparable performance to the EcoVAD teacher model, indicating a promising approach to overcoming computational barriers for real-time ecological monitoring.
Combinatorial Optimization (CO) problems over graphs appear routinely in many applications such as in optimizing traffic, viral marketing in social networks, and matching for job allocation. Due to their combinatorial nature, these problems are often NP-hard. Existing approximation algorithms and heuristics rely on the search space to find the solutions and become time-consuming when this space is large. In this paper, we design a neural method called COMBHelper to reduce this space and thus improve the efficiency of the traditional CO algorithms based on node selection. Specifically, it employs a Graph Neural Network (GNN) to identify promising nodes for the solution set. This pruned search space is then fed to the traditional CO algorithms. COMBHelper also uses a Knowledge Distillation (KD) module and a problem-specific boosting module to bring further efficiency and efficacy. Our extensive experiments show that the traditional CO algorithms with COMBHelper are at least 2 times faster than their original versions.
Knowledge distillation, a technique for model compression and performance enhancement, has gained significant traction in Neural Machine Translation (NMT). However, existing research primarily focuses on empirical applications, and there is a lack of comprehensive understanding of how student model capacity, data complexity, and decoding strategies collectively influence distillation effectiveness. Addressing this gap, our study conducts an in-depth investigation into these factors, particularly focusing on their interplay in word-level and sequence-level distillation within NMT. Through extensive experimentation across datasets like IWSLT13 En$\rightarrow$Fr, IWSLT14 En$\rightarrow$De, and others, we empirically validate hypotheses related to the impact of these factors on knowledge distillation. Our research not only elucidates the significant influence of model capacity, data complexity, and decoding strategies on distillation effectiveness but also introduces a novel, optimized distillation approach. This approach, when applied to the IWSLT14 de$\rightarrow$en translation task, achieves state-of-the-art performance, demonstrating its practical efficacy in advancing the field of NMT.
Polyp segmentation, a contentious issue in medical imaging, has seen numerous proposed methods aimed at improving the quality of segmented masks. While current state-of-the-art techniques yield impressive results, the size and computational cost of these models create challenges for practical industry applications. To address this challenge, we present KDAS, a Knowledge Distillation framework that incorporates attention supervision, and our proposed Symmetrical Guiding Module. This framework is designed to facilitate a compact student model with fewer parameters, allowing it to learn the strengths of the teacher model and mitigate the inconsistency between teacher features and student features, a common challenge in Knowledge Distillation, via the Symmetrical Guiding Module. Through extensive experiments, our compact models demonstrate their strength by achieving competitive results with state-of-the-art methods, offering a promising approach to creating compact models with high accuracy for polyp segmentation and in the medical imaging field. The implementation is available on https://github.com/huyquoctrinh/KDAS.
We propose a cooperative training framework for deep neural network architectures that enables the runtime network depths to change to satisfy dynamic computing resource requirements. In our framework, the number of layers participating in computation can be chosen dynamically to meet performance-cost trade-offs at inference runtime. Our method trains two Teammate nets and a Leader net, and two sets of Teammate sub-networks with various depths through knowledge distillation. The Teammate nets derive sub-networks and transfer knowledge to them, and to each other, while the Leader net guides Teammate nets to ensure accuracy. The approach trains the framework atomically at once instead of individually training various sizes of models; in a sense, the various-sized networks are all trained at once, in a "package deal." The proposed framework is not tied to any specific architecture but can incorporate any existing models/architectures, therefore it can maintain stable results and is insensitive to the size of a dataset's feature map. Compared with other related approaches, it provides comparable accuracy to its full network while various sizes of models are available.
Deep learning-based video quality assessment (deep VQA) has demonstrated significant potential in surpassing conventional metrics, with promising improvements in terms of correlation with human perception. However, the practical deployment of such deep VQA models is often limited due to their high computational complexity and large memory requirements. To address this issue, we aim to significantly reduce the model size and runtime of one of the state-of-the-art deep VQA methods, RankDVQA, by employing a two-phase workflow that integrates pruning-driven model compression with multi-level knowledge distillation. The resulting lightweight full reference quality metric, RankDVQA-mini, requires less than 10% of the model parameters compared to its full version (14% in terms of FLOPs), while still retaining a quality prediction performance that is superior to most existing deep VQA methods. The source code of the RankDVQA-mini has been released at https://chenfeng-bristol.github.io/RankDVQA-mini/ for public evaluation.
Knowledge Distillation (KD) emerges as one of the most promising compression technologies to run advanced deep neural networks on resource-limited devices. In order to train a small network (student) under the guidance of a large network (teacher), the intuitive method is regularizing the feature maps or logits of the student using the teacher's information. However, existing methods either over-restrict the student to learn all information from the teacher, which lead to some bad local minimum, or use various fancy and elaborate modules to process and align features, which are complex and lack generality. In this work, we proposed an abstract and general paradigm for the KD task, referred to as DIMensionality Reduction KD (RdimKD), which solely relies on dimensionality reduction, with a very minor modification to naive L2 loss. RdimKD straightforwardly utilizes a projection matrix to project both the teacher's and student's feature maps onto a low-dimensional subspace, which are then optimized during training. RdimKD achieves the goal in the simplest way that not only does the student get valuable information from the teacher, but it also ensures sufficient flexibility to adapt to the student's low-capacity reality. Our extensive empirical findings indicate the effectiveness of RdimKD across various learning tasks and diverse network architectures.
In this paper, we attempt to specialize the VLM model for OWOD tasks by distilling its open-world knowledge into a language-agnostic detector. Surprisingly, we observe that the combination of a simple \textbf{knowledge distillation} approach and the automatic pseudo-labeling mechanism in OWOD can achieve better performance for unknown object detection, even with a small amount of data. Unfortunately, knowledge distillation for unknown objects severely affects the learning of detectors with conventional structures for known objects, leading to catastrophic forgetting. To alleviate these problems, we propose the \textbf{down-weight loss function} for knowledge distillation from vision-language to single vision modality. Meanwhile, we propose the \textbf{cascade decouple decoding structure} that decouples the learning of localization and recognition to reduce the impact of category interactions of known and unknown objects on the localization learning process. Ablation experiments demonstrate that both of them are effective in mitigating the impact of open-world knowledge distillation on the learning of known objects. Additionally, to alleviate the current lack of comprehensive benchmarks for evaluating the ability of the open-world detector to detect unknown objects in the open world, we propose two benchmarks, which we name "\textbf{StandardSet}$\heartsuit$" and "\textbf{IntensiveSet}$\spadesuit$" respectively, based on the complexity of their testing scenarios. Comprehensive experiments performed on OWOD, MS-COCO, and our proposed benchmarks demonstrate the effectiveness of our methods. The code and proposed dataset are available at \url{https://github.com/xiaomabufei/SKDF}.
Federated learning (FL) provides a decentralized machine learning paradigm where a server collaborates with a group of clients to learn a global model without accessing the clients' data. User heterogeneity is a significant challenge for FL, which together with the class-distribution imbalance further enhances the difficulty of FL. Great progress has been made in large vision-language models, such as Contrastive Language-Image Pre-training (CLIP), which paves a new way for image classification and object recognition. Inspired by the success of CLIP on few-shot and zero-shot learning, we use CLIP to optimize the federated learning between server and client models under its vision-language supervision. It is promising to mitigate the user heterogeneity and class-distribution balance due to the powerful cross-modality representation and rich open-vocabulary prior knowledge. In this paper, we propose the CLIP-guided FL (CLIP2FL) method on heterogeneous and long-tailed data. In CLIP2FL, the knowledge of the off-the-shelf CLIP model is transferred to the client-server models, and a bridge is built between the client and server. Specifically, for client-side learning, knowledge distillation is conducted between client models and CLIP to improve the ability of client-side feature representation. For server-side learning, in order to mitigate the heterogeneity and class-distribution imbalance, we generate federated features to retrain the server model. A prototype contrastive learning with the supervision of the text encoder of CLIP is introduced to generate federated features depending on the client-side gradients, and they are used to retrain a balanced server classifier.
Knowledge distillation (KD), a technique widely employed in computer vision, has emerged as a de facto standard for improving the performance of small neural networks. However, prevailing KD-based approaches in video tasks primarily focus on designing loss functions and fusing cross-modal information. This overlooks the spatial-temporal feature semantics, resulting in limited advancements in model compression. Addressing this gap, our paper introduces an innovative knowledge distillation framework, with the generative model for training a lightweight student model. In particular, the framework is organized into two steps: the initial phase is Feature Representation, wherein a generative model-based attention module is trained to represent feature semantics; Subsequently, the Generative-based Feature Distillation phase encompasses both Generative Distillation and Attention Distillation, with the objective of transferring attention-based feature semantics with the generative model. The efficacy of our approach is demonstrated through comprehensive experiments on diverse popular datasets, proving considerable enhancements in video action recognition task. Moreover, the effectiveness of our proposed framework is validated in the context of more intricate video action detection task. Our code is available at https://github.com/aaai-24/Generative-based-KD.
Nighttime unmanned aerial vehicle (UAV) tracking has been facilitated with indispensable plug-and-play low-light enhancers. However, the introduction of low-light enhancers increases the extra computational burden for the UAV, significantly hindering the development of real-time UAV applications. Meanwhile, these state-of-the-art (SOTA) enhancers lack tight coupling with the advanced daytime UAV tracking approach. To solve the above issues, this work proposes a novel mutual-learning knowledge distillation framework for nighttime UAV tracking, i.e., MLKD. This framework is constructed to learn a compact and fast nighttime tracker via knowledge transferring from the teacher and knowledge sharing among various students. Specifically, an advanced teacher based on a SOTA enhancer and a superior tracking backbone is adopted for guiding the student based only on the tight coupling-aware tracking backbone to directly extract nighttime object features. To address the biased learning of a single student, diverse lightweight students with different distillation methods are constructed to focus on various aspects of the teacher's knowledge. Moreover, an innovative mutual-learning room is designed to elect the superior student candidate to assist the remaining students frame-by-frame in the training phase. Furthermore, the final best student, i.e., MLKD-Track, is selected through the testing dataset. Extensive experiments demonstrate the effectiveness and superiority of MLKD and MLKD-Track. The practicality of the MLKD-Track is verified in real-world tests with different challenging situations. The code is available at https://github.com/lyfeng001/MLKD.
Efficient traffic signal control is critical for reducing traffic congestion and improving overall transportation efficiency. The dynamic nature of traffic flow has prompted researchers to explore Reinforcement Learning (RL) for traffic signal control (TSC). Compared with traditional methods, RL-based solutions have shown preferable performance. However, the application of RL-based traffic signal controllers in the real world is limited by the low sample efficiency and high computational requirements of these solutions. In this work, we propose DTLight, a simple yet powerful lightweight Decision Transformer-based TSC method that can learn policy from easily accessible offline datasets. DTLight novelly leverages knowledge distillation to learn a lightweight controller from a well-trained larger teacher model to reduce implementation computation. Additionally, it integrates adapter modules to mitigate the expenses associated with fine-tuning, which makes DTLight practical for online adaptation with minimal computation and only a few fine-tuning steps during real deployment. Moreover, DTLight is further enhanced to be more applicable to real-world TSC problems. Extensive experiments on synthetic and real-world scenarios show that DTLight pre-trained purely on offline datasets can outperform state-of-the-art online RL-based methods in most scenarios. Experiment results also show that online fine-tuning further improves the performance of DTLight by up to 42.6% over the best online RL baseline methods. In this work, we also introduce Datasets specifically designed for TSC with offline RL (referred to as DTRL). Our datasets and code are publicly available.
We present NovaCOMET, an open commonsense knowledge model, that combines the best aspects of knowledge and general task models. Compared to previous knowledge models, NovaCOMET allows open-format relations enabling direct application to reasoning tasks; compared to general task models like Flan-T5, it explicitly centers knowledge, enabling superior performance for commonsense reasoning. NovaCOMET leverages the knowledge of opaque proprietary models to create an open knowledge pipeline. First, knowledge is symbolically distilled into NovATOMIC, a publicly-released discrete knowledge graph which can be audited, critiqued, and filtered. Next, we train NovaCOMET on NovATOMIC by fine-tuning an open-source pretrained model. NovaCOMET uses an open-format training objective, replacing the fixed relation sets of past knowledge models, enabling arbitrary structures within the data to serve as inputs or outputs. The resulting generation model, optionally augmented with human annotation, matches or exceeds comparable open task models like Flan-T5 on a range of commonsense generation tasks. NovaCOMET serves as a counterexample to the contemporary focus on instruction tuning only, demonstrating a distinct advantage to explicitly modeling commonsense knowledge as well.
Large Language Models (LLMs) trained with self-supervision on vast corpora of web text fit to the social biases of that text. Without intervention, these social biases persist in the model's predictions in downstream tasks, leading to representational harm. Many strategies have been proposed to mitigate the effects of inappropriate social biases learned during pretraining. Simultaneously, methods for model compression have become increasingly popular to reduce the computational burden of LLMs. Despite the popularity and need for both approaches, little work has been done to explore the interplay between these two. We perform a carefully controlled study of the impact of model compression via quantization and knowledge distillation on measures of social bias in LLMs. Longer pretraining and larger models led to higher social bias, and quantization showed a regularizer effect with its best trade-off around 20% of the original pretraining time.
Adversarial Training (AT) has been widely proved to be an effective method to improve the adversarial robustness against adversarial examples for Deep Neural Networks (DNNs). As a variant of AT, Adversarial Robustness Distillation (ARD) has demonstrated its superior performance in improving the robustness of small student models with the guidance of large teacher models. However, both AT and ARD encounter the robust fairness problem: these models exhibit strong robustness when facing part of classes (easy class), but weak robustness when facing others (hard class). In this paper, we give an in-depth analysis of the potential factors and argue that the smoothness degree of samples' soft labels for different classes (i.e., hard class or easy class) will affect the robust fairness of DNNs from both empirical observation and theoretical analysis. Based on the above finding, we propose an Anti-Bias Soft Label Distillation (ABSLD) method to mitigate the adversarial robust fairness problem within the framework of Knowledge Distillation (KD). Specifically, ABSLD adaptively reduces the student's error risk gap between different classes to achieve fairness by adjusting the class-wise smoothness degree of samples' soft labels during the training process, and the smoothness degree of soft labels is controlled by assigning different temperatures in KD to different classes. Extensive experiments demonstrate that ABSLD outperforms state-of-the-art AT, ARD, and robust fairness methods in the comprehensive metric (Normalized Standard Deviation) of robustness and fairness.
Many deep learning based automated medical image segmentation systems, in reality, face difficulties in deployment due to the cost of massive data annotation and high latency in model iteration. We propose a dynamic interactive learning framework that addresses these challenges by integrating interactive segmentation into end-to-end weak supervised learning with streaming tasks. We develop novel replay and label smoothing schemes that overcome catastrophic forgetting and improve online learning robustness. For each image, our multi-round interactive segmentation module simultaneously optimizes both front-end predictions and deep learning segmenter. In each round, a 3D "proxy mask" is propagated from sparse user inputs based on image registration, serving as weak supervision that enable knowledge distillation from the unknown ground truth. In return, the trained segmenter explicitly guides next step's user interventions according to a spatial residual map from consecutive front or back-end predictions. Evaluation on 3D segmentation tasks (NCI-ISBI2013 and BraTS2015) shows that our framework generates online learning performances that match offline training benchmark. In addition, with a 62% reduction in total annotation efforts, our framework produces competitive dice scores comparing to online and offline learning which equipped with full ground truth. Furthermore, such a framework, with its flexibility and responsiveness, could be deployed behind hospital firewall that guarantees data security and easy maintenance.
In federated learning (FL), data heterogeneity is one key bottleneck that causes model divergence and limits performance. Addressing this, existing methods often regard data heterogeneity as an inherent property and propose to mitigate its adverse effects by correcting models. In this paper, we seek to break this inherent property by generating data to complement the original dataset to fundamentally mitigate heterogeneity level. As a novel attempt from the perspective of data, we propose federated learning with consensus-oriented generation (FedCOG). FedCOG consists of two key components at the client side: complementary data generation, which generates data extracted from the shared global model to complement the original dataset, and knowledge-distillation-based model training, which distills knowledge from global model to local model based on the generated data to mitigate over-fitting the original heterogeneous dataset. FedCOG has two critical advantages: 1) it can be a plug-and-play module to further improve the performance of most existing FL methods, and 2) it is naturally compatible with standard FL protocols such as Secure Aggregation since it makes no modification in communication process. Extensive experiments on classical and real-world FL datasets show that FedCOG consistently outperforms state-of-the-art methods.
Neural radiance fields (NeRF) is a promising approach for generating photorealistic images and representing complex scenes. However, when processing data sequentially, it can suffer from catastrophic forgetting, where previous data is easily forgotten after training with new data. Existing incremental learning methods using knowledge distillation assume that continuous data chunks contain both 2D images and corresponding camera pose parameters, pre-estimated from the complete dataset. This poses a paradox as the necessary camera pose must be estimated from the entire dataset, even though the data arrives sequentially and future chunks are inaccessible. In contrast, we focus on a practical scenario where camera poses are unknown. We propose IL-NeRF, a novel framework for incremental NeRF training, to address this challenge. IL-NeRF's key idea lies in selecting a set of past camera poses as references to initialize and align the camera poses of incoming image data. This is followed by a joint optimization of camera poses and replay-based NeRF distillation. Our experiments on real-world indoor and outdoor scenes show that IL-NeRF handles incremental NeRF training and outperforms the baselines by up to $54.04\%$ in rendering quality.
Instruction following vision-language (VL) models offer a flexible interface that supports a broad range of multimodal tasks in a zero-shot fashion. However, interfaces that operate on full images do not directly enable the user to "point to" and access specific regions within images. This capability is important not only to support reference-grounded VL benchmarks, but also, for practical applications that require precise within-image reasoning. We build Localized Visual Commonsense models, which allow users to specify (multiple) regions as input. We train our model by sampling localized commonsense knowledge from a large language model (LLM): specifically, we prompt an LLM to collect commonsense knowledge given a global literal image description and a local literal region description automatically generated by a set of VL models. With a separately trained critic model that selects high-quality examples, we find that training on the localized commonsense corpus can successfully distill existing VL models to support a reference-as-input interface. Empirical results and human evaluations in a zero-shot setup demonstrate that our distillation method results in more precise VL models of reasoning compared to a baseline of passing a generated referring expression to an LLM.
We propose a conceptually simple and lightweight framework for improving the robustness of vision models through the combination of knowledge distillation and data augmentation. We address the conjecture that larger models do not make for better teachers by showing strong gains in out-of-distribution robustness when distilling from pretrained foundation models. Following this finding, we propose Discrete Adversarial Distillation (DAD), which leverages a robust teacher to generate adversarial examples and a VQGAN to discretize them, creating more informative samples than standard data augmentation techniques. We provide a theoretical framework for the use of a robust teacher in the knowledge distillation with data augmentation setting and demonstrate strong gains in out-of-distribution robustness and clean accuracy across different student architectures. Notably, our method adds minor computational overhead compared to similar techniques and can be easily combined with other data augmentations for further improvements.