Deep neural networks have seen enormous success in various real-world applications. Beyond their predictions as point estimates, increasing attention has been focused on quantifying the uncertainty of their predictions. In this review, we show that the uncertainty of deep neural networks is not only important in a sense of interpretability and transparency, but also crucial in further advancing their performance, particularly in learning systems seeking robustness and efficiency. We will generalize the definition of the uncertainty of deep neural networks to any number or vector that is associated with an input or an input-label pair, and catalog existing methods on ``mining'' such uncertainty from a deep model. We will include those methods from the classic field of uncertainty quantification as well as those methods that are specific to deep neural networks. We then show a wide spectrum of applications of such generalized uncertainty in realistic learning tasks including robust learning such as noisy learning, adversarially robust learning; data-efficient learning such as semi-supervised and weakly-supervised learning; and model-efficient learning such as model compression and knowledge distillation.
The use of Neural Architecture Search (NAS) techniques to automate the design of neural networks has become increasingly popular in recent years. The proliferation of devices with different hardware characteristics using such neural networks, as well as the need to reduce the power consumption for their search, has led to the realisation of Once-For-All (OFA), an eco-friendly algorithm characterised by the ability to generate easily adaptable models through a single learning process. In order to improve this paradigm and develop high-performance yet eco-friendly NAS techniques, this paper presents OFAv2, the extension of OFA aimed at improving its performance while maintaining the same ecological advantage. The algorithm is improved from an architectural point of view by including early exits, parallel blocks and dense skip connections. The training process is extended by two new phases called Elastic Level and Elastic Height. A new Knowledge Distillation technique is presented to handle multi-output networks, and finally a new strategy for dynamic teacher network selection is proposed. These modifications allow OFAv2 to improve its accuracy performance on the Tiny ImageNet dataset by up to 12.07% compared to the original version of OFA, while maintaining the algorithm flexibility and advantages.
Graph Neural Networks (GNNs) have attracted tremendous attention by demonstrating their capability to handle graph data. However, they are difficult to be deployed in resource-limited devices due to model sizes and scalability constraints imposed by the multi-hop data dependency. In addition, real-world graphs usually possess complex structural information and features. Therefore, to improve the applicability of GNNs and fully encode the complicated topological information, knowledge distillation on graphs (KDG) has been introduced to build a smaller yet effective model and exploit more knowledge from data, leading to model compression and performance improvement. Recently, KDG has achieved considerable progress with many studies proposed. In this survey, we systematically review these works. Specifically, we first introduce KDG challenges and bases, then categorize and summarize existing works of KDG by answering the following three questions: 1) what to distillate, 2) who to whom, and 3) how to distillate. Finally, we share our thoughts on future research directions.
Deep learning empowers the mainstream medical image segmentation methods. Nevertheless current deep segmentation approaches are not capable of efficiently and effectively adapting and updating the trained models when new incremental segmentation classes (along with new training datasets or not) are required to be added. In real clinical environment, it can be preferred that segmentation models could be dynamically extended to segment new organs/tumors without the (re-)access to previous training datasets due to obstacles of patient privacy and data storage. This process can be viewed as a continual semantic segmentation (CSS) problem, being understudied for multi-organ segmentation. In this work, we propose a new architectural CSS learning framework to learn a single deep segmentation model for segmenting a total of 143 whole-body organs. Using the encoder/decoder network structure, we demonstrate that a continually-trained then frozen encoder coupled with incrementally-added decoders can extract and preserve sufficiently representative image features for new classes to be subsequently and validly segmented. To maintain a single network model complexity, we trim each decoder progressively using neural architecture search and teacher-student based knowledge distillation. To incorporate with both healthy and pathological organs appearing in different datasets, a novel anomaly-aware and confidence learning module is proposed to merge the overlapped organ predictions, originated from different decoders. Trained and validated on 3D CT scans of 2500+ patients from four datasets, our single network can segment total 143 whole-body organs with very high accuracy, closely reaching the upper bound performance level by training four separate segmentation models (i.e., one model per dataset/task).
Knowledge distillation addresses the problem of transferring knowledge from a teacher model to a student model. In this process, we typically have multiple types of knowledge extracted from the teacher model. The problem is to make full use of them to train the student model. Our preliminary study shows that: (1) not all of the knowledge is necessary for learning a good student model, and (2) knowledge distillation can benefit from certain knowledge at different training steps. In response to these, we propose an actor-critic approach to selecting appropriate knowledge to transfer during the process of knowledge distillation. In addition, we offer a refinement of the training algorithm to ease the computational burden. Experimental results on the GLUE datasets show that our method outperforms several strong knowledge distillation baselines significantly.
As a general model compression paradigm, feature-based knowledge distillation allows the student model to learn expressive features from the teacher counterpart. In this paper, we mainly focus on designing an effective feature-distillation framework and propose a spatial-channel adaptive masked distillation (AMD) network for object detection. More specifically, in order to accurately reconstruct important feature regions, we first perform attention-guided feature masking on the feature map of the student network, such that we can identify the important features via spatially adaptive feature masking instead of random masking in the previous methods. In addition, we employ a simple and efficient module to allow the student network channel to be adaptive, improving its model capability in object perception and detection. In contrast to the previous methods, more crucial object-aware features can be reconstructed and learned from the proposed network, which is conducive to accurate object detection. The empirical experiments demonstrate the superiority of our method: with the help of our proposed distillation method, the student networks report 41.3%, 42.4%, and 42.7% mAP scores when RetinaNet, Cascade Mask-RCNN and RepPoints are respectively used as the teacher framework for object detection, which outperforms the previous state-of-the-art distillation methods including FGD and MGD.
Large neural models (such as Transformers) achieve state-of-the-art performance for information retrieval (IR). In this paper, we aim to improve distillation methods that pave the way for the resource-efficient deployment of such models in practice. Inspired by our theoretical analysis of the teacher-student generalization gap for IR models, we propose a novel distillation approach that leverages the relative geometry among queries and documents learned by the large teacher model. Unlike existing teacher score-based distillation methods, our proposed approach employs embedding matching tasks to provide a stronger signal to align the representations of the teacher and student models. In addition, it utilizes query generation to explore the data manifold to reduce the discrepancies between the student and the teacher where training data is sparse. Furthermore, our analysis also motivates novel asymmetric architectures for student models which realizes better embedding alignment without increasing online inference cost. On standard benchmarks like MSMARCO, we show that our approach successfully distills from both dual-encoder (DE) and cross-encoder (CE) teacher models to 1/10th size asymmetric students that can retain 95-97% of the teacher performance.
Originally proposed as a method for knowledge transfer from one model to another, some recent studies have suggested that knowledge distillation (KD) is in fact a form of regularization. Perhaps the strongest argument of all for this new perspective comes from its apparent similarities with label smoothing (LS). Here we re-examine this stated equivalence between the two methods by comparing the predictive confidences of the models they train. Experiments on four text classification tasks involving models of different sizes show that: (a) In most settings, KD and LS drive model confidence in completely opposite directions, and (b) In KD, the student inherits not only its knowledge but also its confidence from the teacher, reinforcing the classical knowledge transfer view.
Despite the popularity and efficacy of knowledge distillation, there is limited understanding of why it helps. In order to study the generalization behavior of a distilled student, we propose a new theoretical framework that leverages supervision complexity: a measure of alignment between teacher-provided supervision and the student's neural tangent kernel. The framework highlights a delicate interplay among the teacher's accuracy, the student's margin with respect to the teacher predictions, and the complexity of the teacher predictions. Specifically, it provides a rigorous justification for the utility of various techniques that are prevalent in the context of distillation, such as early stopping and temperature scaling. Our analysis further suggests the use of online distillation, where a student receives increasingly more complex supervision from teachers in different stages of their training. We demonstrate efficacy of online distillation and validate the theoretical findings on a range of image classification benchmarks and model architectures.
Although industrial anomaly detection (AD) technology has made significant progress in recent years, generating realistic anomalies and learning priors of normal remain challenging tasks. In this study, we propose an end-to-end industrial anomaly detection method called FractalAD. Training samples are obtained by synthesizing fractal images and patches from normal samples. This fractal anomaly generation method is designed to sample the full morphology of anomalies. Moreover, we designed a backbone knowledge distillation structure to extract prior knowledge contained in normal samples. The differences between a teacher and a student model are converted into anomaly attention using a cosine similarity attention module. The proposed method enables an end-to-end semantic segmentation network to be used for anomaly detection without adding any trainable parameters to the backbone and segmentation head, and has obvious advantages over other methods in training and inference speed.. The results of ablation studies confirmed the effectiveness of fractal anomaly generation and backbone knowledge distillation. The results of performance experiments showed that FractalAD achieved competitive results on the MVTec AD dataset and MVTec 3D-AD dataset compared with other state-of-the-art anomaly detection methods.
Face image translation has made notable progress in recent years. However, when training on limited data, the performance of existing approaches significantly declines. Although some studies have attempted to tackle this problem, they either failed to achieve the few-shot setting (less than 10) or can only get suboptimal results. In this paper, we propose GAN Prior Distillation (GPD) to enable effective few-shot face image translation. GPD contains two models: a teacher network with GAN Prior and a student network that fulfills end-to-end translation. Specifically, we adapt the teacher network trained on large-scale data in the source domain to the target domain with only a few samples, where it can learn the target domain's knowledge. Then, we can achieve few-shot augmentation by generating source domain and target domain images simultaneously with the same latent codes. We propose an anchor-based knowledge distillation module that can fully use the difference between the training and the augmented data to distill the knowledge of the teacher network into the student network. The trained student network achieves excellent generalization performance with the absorption of additional knowledge. Qualitative and quantitative experiments demonstrate that our method achieves superior results than state-of-the-art approaches in a few-shot setting.
Knowledge distillation (KD) is one of the prominent techniques for model compression. In this method, the knowledge of a large network (teacher) is distilled into a model (student) with usually significantly fewer parameters. KD tries to better-match the output of the student model to that of the teacher model based on the knowledge extracts from the forward pass of the teacher network. Although conventional KD is effective for matching the two networks over the given data points, there is no guarantee that these models would match in other areas for which we do not have enough training samples. In this work, we address that problem by generating new auxiliary training samples based on extracting knowledge from the backward pass of the teacher in the areas where the student diverges greatly from the teacher. We compute the difference between the teacher and the student and generate new data samples that maximize the divergence. This is done by perturbing data samples in the direction of the gradient of the difference between the student and the teacher. Augmenting the training set by adding this auxiliary improves the performance of KD significantly and leads to a closer match between the student and the teacher. Using this approach, when data samples come from a discrete domain, such as applications of natural language processing (NLP) and language understanding, is not trivial. However, we show how this technique can be used successfully in such applications. We evaluated the performance of our method on various tasks in computer vision and NLP domains and got promising results.
The widespread emergence of smart devices for ECG has sparked demand for intelligent single-lead ECG-based diagnostic systems. However, it is challenging to develop a single-lead-based ECG interpretation model for multiple diseases diagnosis due to the lack of some key disease information. In this work, we propose inter-lead Multi-View Knowledge Transferring of ECG (MVKT-ECG) to boost single-lead ECG's ability for multi-label disease diagnosis. This training strategy can transfer superior disease knowledge from multiple different views of ECG (e.g. 12-lead ECG) to single-lead-based ECG interpretation model to mine details in single-lead ECG signals that are easily overlooked by neural networks. MVKT-ECG allows this lead variety as a supervision signal within a teacher-student paradigm, where the teacher observes multi-lead ECG educates a student who observes only single-lead ECG. Since the mutual disease information between the single-lead ECG and muli-lead ECG plays a key role in knowledge transferring, we present a new disease-aware Contrastive Lead-information Transferring(CLT) to improve the mutual disease information between the single-lead ECG and muli-lead ECG. Moreover, We modify traditional Knowledge Distillation to multi-label disease Knowledge Distillation (MKD) to make it applicable for multi-label disease diagnosis. The comprehensive experiments verify that MVKT-ECG has an excellent performance in improving the diagnostic effect of single-lead ECG.
Large-scale pre-trained language models (PLMs) have shown great potential in natural language processing tasks. Leveraging the capabilities of PLMs to enhance automatic speech recognition (ASR) systems has also emerged as a promising research direction. However, previous works may be limited by the inflexible structures of PLMs and the insufficient utilization of PLMs. To alleviate these problems, we propose the hierarchical knowledge distillation (HKD) on the continuous integrate-and-fire (CIF) based ASR models. To transfer knowledge from PLMs to the ASR models, HKD employs cross-modal knowledge distillation with contrastive loss at the acoustic level and knowledge distillation with regression loss at the linguistic level. Compared with the original CIF-based model, our method achieves 15% and 9% relative error rate reduction on the AISHELL-1 and LibriSpeech datasets, respectively.
Knowledge distillation (KD) has been widely used to improve the test accuracy of a "student" network, by training it to mimic the soft probabilities of a trained "teacher" network. Yet, it has been shown in recent work that, despite being trained to fit the teacher's probabilities, the student may not only significantly deviate from the teacher probabilities, but may also outdo than the teacher in performance. Our work aims to reconcile this seemingly paradoxical observation. Specifically, we characterize the precise nature of the student-teacher deviations, and argue how they can co-occur with better generalization. First, through experiments on image and language data, we identify that these probability deviations correspond to the student systematically exaggerating the confidence levels of the teacher. Next, we theoretically and empirically establish another form of exaggeration in some simple settings: KD exaggerates the implicit bias of gradient descent in converging faster along the top eigendirections of the data. Finally, we tie these two observations together: we demonstrate that the exaggerated bias of KD can simultaneously result in both (a) the exaggeration of confidence and (b) the improved generalization of the student, thus offering a resolution to the apparent paradox. Our analysis brings existing theory and practice closer by considering the role of gradient descent in KD and by demonstrating the exaggerated bias effect in both theoretical and empirical settings.
In this article, we use probing to investigate phenomena that occur during fine-tuning and knowledge distillation of a BERT-based natural language understanding (NLU) model. Our ultimate purpose was to use probing to better understand practical production problems and consequently to build better NLU models. We designed experiments to see how fine-tuning changes the linguistic capabilities of BERT, what the optimal size of the fine-tuning dataset is, and what amount of information is contained in a distilled NLU based on a tiny Transformer. The results of the experiments show that the probing paradigm in its current form is not well suited to answer such questions. Structural, Edge and Conditional probes do not take into account how easy it is to decode probed information. Consequently, we conclude that quantification of information decodability is critical for many practical applications of the probing paradigm.
Modeling text-based time-series to make prediction about a future event or outcome is an important task with a wide range of applications. The standard approach is to train and test the model using the same input window, but this approach neglects the data collected in longer input windows between the prediction time and the final outcome, which are often available during training. In this study, we propose to treat this neglected text as privileged information available during training to enhance early prediction modeling through knowledge distillation, presented as Learning using Privileged tIme-sEries Text (LuPIET). We evaluate the method on clinical and social media text, with four clinical prediction tasks based on clinical notes and two mental health prediction tasks based on social media posts. Our results show LuPIET is effective in enhancing text-based early predictions, though one may need to consider choosing the appropriate text representation and windows for privileged text to achieve optimal performance. Compared to two other methods using transfer learning and mixed training, LuPIET offers more stable improvements over the baseline, standard training. As far as we are concerned, this is the first study to examine learning using privileged information for time-series in the NLP context.
Heterogeneity of data distributed across clients limits the performance of global models trained through federated learning, especially in the settings with highly imbalanced class distributions of local datasets. In recent years, personalized federated learning (pFL) has emerged as a potential solution to the challenges presented by heterogeneous data. However, existing pFL methods typically enhance performance of local models at the expense of the global model's accuracy. We propose FedHKD (Federated Hyper-Knowledge Distillation), a novel FL algorithm in which clients rely on knowledge distillation (KD) to train local models. In particular, each client extracts and sends to the server the means of local data representations and the corresponding soft predictions -- information that we refer to as ``hyper-knowledge". The server aggregates this information and broadcasts it to the clients in support of local training. Notably, unlike other KD-based pFL methods, FedHKD does not rely on a public dataset nor it deploys a generative model at the server. We analyze convergence of FedHKD and conduct extensive experiments on visual datasets in a variety of scenarios, demonstrating that FedHKD provides significant improvement in both personalized as well as global model performance compared to state-of-the-art FL methods designed for heterogeneous data settings.
In this paper, we consider the problem of simultaneously detecting objects and inferring their visual attributes in an image, even for those with no manual annotations provided at the training stage, resembling an open-vocabulary scenario. To achieve this goal, we make the following contributions: (i) we start with a naive two-stage approach for open-vocabulary object detection and attribute classification, termed CLIP-Attr. The candidate objects are first proposed with an offline RPN and later classified for semantic category and attributes; (ii) we combine all available datasets and train with a federated strategy to finetune the CLIP model, aligning the visual representation with attributes, additionally, we investigate the efficacy of leveraging freely available online image-caption pairs under weakly supervised learning; (iii) in pursuit of efficiency, we train a Faster-RCNN type model end-to-end with knowledge distillation, that performs class-agnostic object proposals and classification on semantic categories and attributes with classifiers generated from a text encoder; Finally, (iv) we conduct extensive experiments on VAW, MS-COCO, LSA, and OVAD datasets, and show that recognition of semantic category and attributes is complementary for visual scene understanding, i.e., jointly training object detection and attributes prediction largely outperform existing approaches that treat the two tasks independently, demonstrating strong generalization ability to novel attributes and categories.
Self-supervised methods in vision have been mostly focused on large architectures as they seem to suffer from a significant performance drop for smaller architectures. In this paper, we propose a simple self-supervised distillation technique that can train high performance low-compute neural networks. Our main insight is that existing joint-embedding based SSL methods can be repurposed for knowledge distillation from a large self-supervised teacher to a small student model. Thus, we call our method Replace one Branch (RoB) as it simply replaces one branch of the joint-embedding training with a large teacher model. RoB is widely applicable to a number of architectures such as small ResNets, MobileNets and ViT, and pretrained models such as DINO, SwAV or iBOT. When pretraining on the ImageNet dataset, RoB yields models that compete with supervised knowledge distillation. When applied to MSN, RoB produces students with strong semi-supervised capabilities. Finally, our best ViT-Tiny models improve over prior SSL state-of-the-art on ImageNet by $2.3\%$ and are on par or better than a supervised distilled DeiT on five downstream transfer tasks (iNaturalist, CIFAR, Clevr/Count, Clevr/Dist and Places). We hope RoB enables practical self-supervision at smaller scale.