The rising of deep learning has facilitated the development of single image super-resolution (SISR). However, the growing burdensome model complexity and memory occupation severely hinder its practical deployments on resource-limited devices. In this paper, we propose a novel joint-distillation (JDSR) framework to boost the representation of various off-the-shelf lightweight SR models. The framework includes two stages: the superior LR generation and the joint-distillation learning. The superior LR is obtained from the HR image itself. With less than $300$K parameters, the peer network using superior LR as input can achieve comparable SR performance with large models, e.g., RCAN, with 15M parameters, which enables it as the input of peer network to save the training expense. The joint-distillation learning consists of internal self-distillation and external mutual learning. The internal self-distillation aims to achieve model self-boosting by transferring the knowledge from the deeper SR output to the shallower one. Specifically, each intermediate SR output is supervised by the HR image and the soft label from subsequent deeper outputs. To shrink the capacity gap between shallow and deep layers, a soft label generator is designed in a progressive backward fusion way with meta-learning for adaptive weight fine-tuning. The external mutual learning focuses on obtaining interaction information from a peer network in the process. Moreover, a curriculum learning strategy and a performance gap threshold are introduced for balancing the convergence rate of the original SR model and its peer network. Comprehensive experiments on benchmark datasets demonstrate that our proposal improves the performance of recent lightweight SR models by a large margin, with the same model architecture and inference expense.
The performance of local feature descriptors degrades in the presence of large rotation variations. To address this issue, we present an efficient approach to learning rotation invariant descriptors. Specifically, we propose Rotated Kernel Fusion (RKF) which imposes rotations on the convolution kernel to improve the inherent nature of CNN. Since RKF can be processed by the subsequent re-parameterization, no extra computational costs will be introduced in the inference stage. Moreover, we present Multi-oriented Feature Aggregation (MOFA) which aggregates features extracted from multiple rotated versions of the input image and can provide auxiliary knowledge for the training of RKF by leveraging the distillation strategy. We refer to the distilled RKF model as DRKF. Besides the evaluation on a rotation-augmented version of the public dataset HPatches, we also contribute a new dataset named DiverseBEV which is collected during the drone's flight and consists of bird's eye view images with large viewpoint changes and camera rotations. Extensive experiments show that our method can outperform other state-of-the-art techniques when exposed to large rotation variations.
Data-free Knowledge Distillation (DFKD) has attracted attention recently thanks to its appealing capability of transferring knowledge from a teacher network to a student network without using training data. The main idea is to use a generator to synthesize data for training the student. As the generator gets updated, the distribution of synthetic data will change. Such distribution shift could be large if the generator and the student are trained adversarially, causing the student to forget the knowledge it acquired at previous steps. To alleviate this problem, we propose a simple yet effective method called Momentum Adversarial Distillation (MAD) which maintains an exponential moving average (EMA) copy of the generator and uses synthetic samples from both the generator and the EMA generator to train the student. Since the EMA generator can be considered as an ensemble of the generator's old versions and often undergoes a smaller change in updates compared to the generator, training on its synthetic samples can help the student recall the past knowledge and prevent the student from adapting too quickly to new updates of the generator. Our experiments on six benchmark datasets including big datasets like ImageNet and Places365 demonstrate the superior performance of MAD over competing methods for handling the large distribution shift problem. Our method also compares favorably to existing DFKD methods and even achieves state-of-the-art results in some cases.
Knowledge Distillation (KD) for object detection aims to train a compact detector by transferring knowledge from a teacher model. Since the teacher model perceives data in a way different from humans, existing KD methods only distill knowledge that is consistent with labels annotated by human expert while neglecting knowledge that is not consistent with human perception, which results in insufficient distillation and sub-optimal performance. In this paper, we propose inconsistent knowledge distillation (IKD), which aims to distill knowledge inherent in the teacher model's counter-intuitive perceptions. We start by considering the teacher model's counter-intuitive perceptions of frequency and non-robust features. Unlike previous works that exploit fine-grained features or introduce additional regularizations, we extract inconsistent knowledge by providing diverse input using data augmentation. Specifically, we propose a sample-specific data augmentation to transfer the teacher model's ability in capturing distinct frequency components and suggest an adversarial feature augmentation to extract the teacher model's perceptions of non-robust features in the data. Extensive experiments demonstrate the effectiveness of our method which outperforms state-of-the-art KD baselines on one-stage, two-stage and anchor-free object detectors (at most +1.0 mAP). Our codes will be made available at \url{https://github.com/JWLiang007/IKD.git}.
While transformers and their variant conformers show promising performance in speech recognition, the parameterized property leads to much memory cost during training and inference. Some works use cross-layer weight-sharing to reduce the parameters of the model. However, the inevitable loss of capacity harms the model performance. To address this issue, this paper proposes a parameter-efficient conformer via sharing sparsely-gated experts. Specifically, we use sparsely-gated mixture-of-experts (MoE) to extend the capacity of a conformer block without increasing computation. Then, the parameters of the grouped conformer blocks are shared so that the number of parameters is reduced. Next, to ensure the shared blocks with the flexibility of adapting representations at different levels, we design the MoE routers and normalization individually. Moreover, we use knowledge distillation to further improve the performance. Experimental results show that the proposed model achieves competitive performance with 1/3 of the parameters of the encoder, compared with the full-parameter model.
Wearable sensor-based human activity recognition (HAR) has emerged as a principal research area and is utilized in a variety of applications. Recently, deep learning-based methods have achieved significant improvement in the HAR field with the development of human-computer interaction applications. However, they are limited to operating in a local neighborhood in the process of a standard convolution neural network, and correlations between different sensors on body positions are ignored. In addition, they still face significant challenging problems with performance degradation due to large gaps in the distribution of training and test data, and behavioral differences between subjects. In this work, we propose a novel Transformer-based Adversarial learning framework for human activity recognition using wearable sensors via Self-KnowledgE Distillation (TASKED), that accounts for individual sensor orientations and spatial and temporal features. The proposed method is capable of learning cross-domain embedding feature representations from multiple subjects datasets using adversarial learning and the maximum mean discrepancy (MMD) regularization to align the data distribution over multiple domains. In the proposed method, we adopt the teacher-free self-knowledge distillation to improve the stability of the training procedure and the performance of human activity recognition. Experimental results show that TASKED not only outperforms state-of-the-art methods on the four real-world public HAR datasets (alone or combined) but also improves the subject generalization effectively.
We propose WS-DINO as a novel framework to use weak label information in learning phenotypic representations from high-content fluorescent images of cells. Our model is based on a knowledge distillation approach with a vision transformer backbone (DINO), and we use this as a benchmark model for our study. Using WS-DINO, we fine-tuned with weak label information available in high-content microscopy screens (treatment and compound) and achieve state-of-the-art performance in not-same-compound mechanism of action prediction on the BBBC021 dataset (98%), and not-same-compound-and-batch performance (96%) using the compound as the weak label. Our method bypasses single cell cropping as a pre-processing step, and using self-attention maps we show that the model learns structurally meaningful phenotypic profiles.
Knowledge distillation (KD) is an effective tool for compressing deep classification models for edge devices. However, the performance of KD is affected by the large capacity gap between the teacher and student networks. Recent methods have resorted to a multiple teacher assistant (TA) setting for KD, which sequentially decreases the size of the teacher model to relatively bridge the size gap between these models. This paper proposes a new technique called Curriculum Expert Selection for Knowledge Distillation (CES-KD) to efficiently enhance the learning of a compact student under the capacity gap problem. This technique is built upon the hypothesis that a student network should be guided gradually using stratified teaching curriculum as it learns easy (hard) data samples better and faster from a lower (higher) capacity teacher network. Specifically, our method is a gradual TA-based KD technique that selects a single teacher per input image based on a curriculum driven by the difficulty in classifying the image. In this work, we empirically verify our hypothesis and rigorously experiment with CIFAR-10, CIFAR-100, CINIC-10, and ImageNet datasets and show improved accuracy on VGG-like models, ResNets, and WideResNets architectures.
The field of continual deep learning is an emerging field and a lot of progress has been made. However, concurrently most of the approaches are only tested on the task of image classification, which is not relevant in the field of intelligent vehicles. Only recently approaches for class-incremental semantic segmentation were proposed. However, all of those approaches are based on some form of knowledge distillation. At the moment there are no investigations on replay-based approaches that are commonly used for object recognition in a continual setting. At the same time while unsupervised domain adaption for semantic segmentation gained a lot of traction, investigations regarding domain-incremental learning in an continual setting is not well-studied. Therefore, the goal of our work is to evaluate and adapt established solutions for continual object recognition to the task of semantic segmentation and to provide baseline methods and evaluation protocols for the task of continual semantic segmentation. We firstly introduce evaluation protocols for the class- and domain-incremental segmentation and analyze selected approaches. We show that the nature of the task of semantic segmentation changes which methods are most effective in mitigating forgetting compared to image classification. Especially, in class-incremental learning knowledge distillation proves to be a vital tool, whereas in domain-incremental learning replay methods are the most effective method.
Class-incremental learning for semantic segmentation (CiSS) is presently a highly researched field which aims at updating a semantic segmentation model by sequentially learning new semantic classes. A major challenge in CiSS is overcoming the effects of catastrophic forgetting, which describes the sudden drop of accuracy on previously learned classes after the model is trained on a new set of classes. Despite latest advances in mitigating catastrophic forgetting, the underlying causes of forgetting specifically in CiSS are not well understood. Therefore, in a set of experiments and representational analyses, we demonstrate that the semantic shift of the background class and a bias towards new classes are the major causes of forgetting in CiSS. Furthermore, we show that both causes mostly manifest themselves in deeper classification layers of the network, while the early layers of the model are not affected. Finally, we demonstrate how both causes are effectively mitigated utilizing the information contained in the background, with the help of knowledge distillation and an unbiased cross-entropy loss.
In unsupervised domain adaptation (UDA), directly adapting from the source to the target domain usually suffers significant discrepancies and leads to insufficient alignment. Thus, many UDA works attempt to vanish the domain gap gradually and softly via various intermediate spaces, dubbed domain bridging (DB). However, for dense prediction tasks such as domain adaptive semantic segmentation (DASS), existing solutions have mostly relied on rough style transfer and how to elegantly bridge domains is still under-explored. In this work, we resort to data mixing to establish a deliberated domain bridging (DDB) for DASS, through which the joint distributions of source and target domains are aligned and interacted with each in the intermediate space. At the heart of DDB lies a dual-path domain bridging step for generating two intermediate domains using the coarse-wise and the fine-wise data mixing techniques, alongside a cross-path knowledge distillation step for taking two complementary models trained on generated intermediate samples as 'teachers' to develop a superior 'student' in a multi-teacher distillation manner. These two optimization steps work in an alternating way and reinforce each other to give rise to DDB with strong adaptation power. Extensive experiments on adaptive segmentation tasks with different settings demonstrate that our DDB significantly outperforms state-of-the-art methods. Code is available at https://github.com/xiaoachen98/DDB.git.
In this work, we propose a novel approach for layerwise representation learning of a trained neural network. In particular, we form a Bregman divergence based on the layer's transfer function and construct an extension of the original Bregman PCA formulation by incorporating a mean vector and normalizing the principal directions with respect to the geometry of the local convex function around the mean. This generalization allows exporting the learned representation as a fixed layer with a non-linearity. As an application to knowledge distillation, we cast the learning problem for the student network as predicting the compression coefficients of the teacher's representations, which are passed as the input to the imported layer. Our empirical findings indicate that our approach is substantially more effective for transferring information between networks than typical teacher-student training using the teacher's penultimate layer representations and soft labels.
We present a systematic study of domain generalization (DG) for tiny neural networks. This problem is critical to on-device machine learning applications but has been overlooked in the literature where research has been merely focused on large models. Tiny neural networks have much fewer parameters and lower complexity and therefore should not be trained the same way as their large counterparts for DG applications. By conducting extensive experiments, we find that knowledge distillation (KD), a well-known technique for model compression, is much better for tackling the on-device DG problem than conventional DG methods. Another interesting observation is that the teacher-student gap on out-of-distribution data is bigger than that on in-distribution data, which highlights the capacity mismatch issue as well as the shortcoming of KD. We further propose a method called out-of-distribution knowledge distillation (OKD) where the idea is to teach the student how the teacher handles out-of-distribution data synthesized via disruptive data augmentation. Without adding any extra parameter to the model -- hence keeping the deployment cost unchanged -- OKD significantly improves DG performance for tiny neural networks in a variety of on-device DG scenarios for image and speech applications. We also contribute a scalable approach for synthesizing visual domain shifts, along with a new suite of DG datasets to complement existing testbeds.
Plasticity and stability are needed in class-incremental learning in order to learn from new data while preserving past knowledge. Due to catastrophic forgetting, finding a compromise between these two properties is particularly challenging when no memory buffer is available. Mainstream methods need to store two deep models since they integrate new classes using fine-tuning with knowledge distillation from the previous incremental state. We propose a method which has similar number of parameters but distributes them differently in order to find a better balance between plasticity and stability. Following an approach already deployed by transfer-based incremental methods, we freeze the feature extractor after the initial state. Classes in the oldest incremental states are trained with this frozen extractor to ensure stability. Recent classes are predicted using partially fine-tuned models in order to introduce plasticity. Our proposed plasticity layer can be incorporated to any transfer-based method designed for exemplar-free incremental learning, and we apply it to two such methods. Evaluation is done with three large-scale datasets. Results show that performance gains are obtained in all tested configurations compared to existing methods.
Previous studies have proved that cross-lingual knowledge distillation can significantly improve the performance of pre-trained models for cross-lingual similarity matching tasks. However, the student model needs to be large in this operation. Otherwise, its performance will drop sharply, thus making it impractical to be deployed to memory-limited devices. To address this issue, we delve into cross-lingual knowledge distillation and propose a multi-stage distillation framework for constructing a small-size but high-performance cross-lingual model. In our framework, contrastive learning, bottleneck, and parameter recurrent strategies are combined to prevent performance from being compromised during the compression process. The experimental results demonstrate that our method can compress the size of XLM-R and MiniLM by more than 50\%, while the performance is only reduced by about 1%.
Continual learning aims to rapidly and continually learn the current task from a sequence of tasks. Compared to other kinds of methods, the methods based on experience replay have shown great advantages to overcome catastrophic forgetting. One common limitation of this method is the data imbalance between the previous and current tasks, which would further aggravate forgetting. Moreover, how to effectively address the stability-plasticity dilemma in this setting is also an urgent problem to be solved. In this paper, we overcome these challenges by proposing a novel framework called Meta-learning update via Multi-scale Knowledge Distillation and Data Augmentation (MMKDDA). Specifically, we apply multiscale knowledge distillation to grasp the evolution of long-range and short-range spatial relationships at different feature levels to alleviate the problem of data imbalance. Besides, our method mixes the samples from the episodic memory and current task in the online continual training procedure, thus alleviating the side influence due to the change of probability distribution. Moreover, we optimize our model via the meta-learning update resorting to the number of tasks seen previously, which is helpful to keep a better balance between stability and plasticity. Finally, our experimental evaluation on four benchmark datasets shows the effectiveness of the proposed MMKDDA framework against other popular baselines, and ablation studies are also conducted to further analyze the role of each component in our framework.
Federated Learning (FL) is a machine learning paradigm where local nodes collaboratively train a central model while the training data remains decentralized. Existing FL methods typically share model parameters or employ co-distillation to address the issue of unbalanced data distribution. However, they suffer from communication bottlenecks. More importantly, they risk privacy leakage. In this work, we develop a privacy preserving and communication efficient method in a FL framework with one-shot offline knowledge distillation using unlabeled, cross-domain public data. We propose a quantized and noisy ensemble of local predictions from completely trained local models for stronger privacy guarantees without sacrificing accuracy. Based on extensive experiments on image classification and text classification tasks, we show that our privacy-preserving method outperforms baseline FL algorithms with superior performance in both accuracy and communication efficiency.
Online Knowledge Distillation (OKD) improves the involved models by reciprocally exploiting the difference between teacher and student. Several crucial bottlenecks over the gap between them -- e.g., Why and when does a large gap harm the performance, especially for student? How to quantify the gap between teacher and student? -- have received limited formal study. In this paper, we propose Switchable Online Knowledge Distillation (SwitOKD), to answer these questions. Instead of focusing on the accuracy gap at test phase by the existing arts, the core idea of SwitOKD is to adaptively calibrate the gap at training phase, namely distillation gap, via a switching strategy between two modes -- expert mode (pause the teacher while keep the student learning) and learning mode (restart the teacher). To possess an appropriate distillation gap, we further devise an adaptive switching threshold, which provides a formal criterion as to when to switch to learning mode or expert mode, and thus improves the student's performance. Meanwhile, the teacher benefits from our adaptive switching threshold and keeps basically on a par with other online arts. We further extend SwitOKD to multiple networks with two basis topologies. Finally, extensive experiments and analysis validate the merits of SwitOKD for classification over the state-of-the-arts. Our code is available at https://github.com/hfutqian/SwitOKD.
Continual learning aims to learn a sequence of tasks by leveraging the knowledge acquired in the past in an online-learning manner while being able to perform well on all previous tasks, this ability is crucial to the artificial intelligence (AI) system, hence continual learning is more suitable for most real-word and complex applicative scenarios compared to the traditional learning pattern. However, the current models usually learn a generic representation base on the class label on each task and an effective strategy is selected to avoid catastrophic forgetting. We postulate that selecting the related and useful parts only from the knowledge obtained to perform each task is more effective than utilizing the whole knowledge. Based on this fact, in this paper we propose a new framework, named Selecting Related Knowledge for Online Continual Learning (SRKOCL), which incorporates an additional efficient channel attention mechanism to pick the particular related knowledge for every task. Our model also combines experience replay and knowledge distillation to circumvent the catastrophic forgetting. Finally, extensive experiments are conducted on different benchmarks and the competitive experimental results demonstrate that our proposed SRKOCL is a promised approach against the state-of-the-art.
We explore the use of knowledge distillation (KD) for learning compact and accurate models that enable classification of animal behavior from accelerometry data on wearable devices. To this end, we take a deep and complex convolutional neural network, known as residual neural network (ResNet), as the teacher model. ResNet is specifically designed for multivariate time-series classification. We use ResNet to distill the knowledge of animal behavior classification datasets into soft labels, which consist of the predicted pseudo-probabilities of every class for each datapoint. We then use the soft labels to train our significantly less complex student models, which are based on the gated recurrent unit (GRU) and multilayer perceptron (MLP). The evaluation results using two real-world animal behavior classification datasets show that the classification accuracy of the student GRU-MLP models improves appreciably through KD, approaching that of the teacher ResNet model. To further reduce the computational and memory requirements of performing inference using the student models trained via KD, we utilize dynamic fixed-point quantization (DQ) through an appropriate modification of the computational graph of the considered models. We implement both unquantized and quantized versions of the developed KD-based models on the embedded systems of our purpose-built collar and ear tag devices to classify animal behavior in situ and in real time. Our evaluations corroborate the effectiveness of KD and DQ in improving the accuracy and efficiency of in-situ animal behavior classification.