Previous Knowledge Distillation based efficient image retrieval methods employs a lightweight network as the student model for fast inference. However, the lightweight student model lacks adequate representation capacity for effective knowledge imitation during the most critical early training period, causing final performance degeneration. To tackle this issue, we propose a Capacity Dynamic Distillation framework, which constructs a student model with editable representation capacity. Specifically, the employed student model is initially a heavy model to fruitfully learn distilled knowledge in the early training epochs, and the student model is gradually compressed during the training. To dynamically adjust the model capacity, our dynamic framework inserts a learnable convolutional layer within each residual block in the student model as the channel importance indicator. The indicator is optimized simultaneously by the image retrieval loss and the compression loss, and a retrieval-guided gradient resetting mechanism is proposed to release the gradient conflict. Extensive experiments show that our method has superior inference speed and accuracy, e.g., on the VeRi-776 dataset, given the ResNet101 as a teacher, our method saves 67.13% model parameters and 65.67% FLOPs (around 24.13% and 21.94% higher than state-of-the-arts) without sacrificing accuracy (around 2.11% mAP higher than state-of-the-arts).
Recent advancements in pre-trained vision-language models, such as CLIP, have enabled the segmentation of arbitrary concepts solely from textual inputs, a process commonly referred to as open-vocabulary semantic segmentation (OVS). However, existing OVS techniques confront a fundamental challenge: the trained classifier tends to overfit on the base classes observed during training, resulting in suboptimal generalization performance to unseen classes. To mitigate this issue, recent studies have proposed the use of an additional frozen pre-trained CLIP for classification. Nonetheless, this approach incurs heavy computational overheads as the CLIP vision encoder must be repeatedly forward-passed for each mask, rendering it impractical for real-world applications. To address this challenge, our objective is to develop a fast OVS model that can perform comparably or better without the extra computational burden of the CLIP image encoder during inference. To this end, we propose a core idea of preserving the generalizable representation when fine-tuning on known classes. Specifically, we introduce a text diversification strategy that generates a set of synonyms for each training category, which prevents the learned representation from collapsing onto specific known category names. Additionally, we employ a text-guided knowledge distillation method to preserve the generalizable knowledge of CLIP. Extensive experiments demonstrate that our proposed model achieves robust generalization performance across various datasets. Furthermore, we perform a preliminary exploration of open-vocabulary video segmentation and present a benchmark that can facilitate future open-vocabulary research in the video domain.
We propose Dataset Reinforcement, a strategy to improve a dataset once such that the accuracy of any model architecture trained on the reinforced dataset is improved at no additional training cost for users. We propose a Dataset Reinforcement strategy based on data augmentation and knowledge distillation. Our generic strategy is designed based on extensive analysis across CNN- and transformer-based models and performing large-scale study of distillation with state-of-the-art models with various data augmentations. We create a reinforced version of the ImageNet training dataset, called ImageNet+, as well as reinforced datasets CIFAR-100+, Flowers-102+, and Food-101+. Models trained with ImageNet+ are more accurate, robust, and calibrated, and transfer well to downstream tasks (e.g., segmentation and detection). As an example, the accuracy of ResNet-50 improves by 1.7% on the ImageNet validation set, 3.5% on ImageNetV2, and 10.0% on ImageNet-R. Expected Calibration Error (ECE) on the ImageNet validation set is also reduced by 9.9%. Using this backbone with Mask-RCNN for object detection on MS-COCO, the mean average precision improves by 0.8%. We reach similar gains for MobileNets, ViTs, and Swin-Transformers. For MobileNetV3 and Swin-Tiny, we observe significant improvements on ImageNet-R/A/C of up to 20% improved robustness. Models pretrained on ImageNet+ and fine-tuned on CIFAR-100+, Flowers-102+, and Food-101+, reach up to 3.4% improved accuracy. The code, datasets, and pretrained models are available at https://github.com/apple/ml-dr.
Graph neural networks (GNNs) have shown the power in representation learning over graph-structured user-item interaction data for collaborative filtering (CF) task. However, with their inherently recursive message propagation among neighboring nodes, existing GNN-based CF models may generate indistinguishable and inaccurate user (item) representations due to the over-smoothing and noise effect with low-pass Laplacian smoothing operators. In addition, the recursive information propagation with the stacked aggregators in the entire graph structures may result in poor scalability in practical applications. Motivated by these limitations, we propose a simple and effective collaborative filtering model (SimRec) that marries the power of knowledge distillation and contrastive learning. In SimRec, adaptive transferring knowledge is enabled between the teacher GNN model and a lightweight student network, to not only preserve the global collaborative signals, but also address the over-smoothing issue with representation recalibration. Empirical results on public datasets show that SimRec archives better efficiency while maintaining superior recommendation performance compared with various strong baselines. Our implementations are publicly available at: https://github.com/HKUDS/SimRec.
Knowledge distillation (KD) has been extensively studied in single-label image classification. However, its efficacy for multi-label classification remains relatively unexplored. In this study, we firstly investigate the effectiveness of classical KD techniques, including logit-based and feature-based methods, for multi-label classification. Our findings indicate that the logit-based method is not well-suited for multi-label classification, as the teacher fails to provide inter-category similarity information or regularization effect on student model's training. Moreover, we observe that feature-based methods struggle to convey compact information of multiple labels simultaneously. Given these limitations, we propose that a suitable dark knowledge should incorporate class-wise information and be highly correlated with the final classification results. To address these issues, we introduce a novel distillation method based on Class Activation Maps (CAMs), which is both effective and straightforward to implement. Across a wide range of settings, CAMs-based distillation consistently outperforms other methods.
Cross-resolution face recognition has become a challenging problem for modern deep face recognition systems. It aims at matching a low-resolution probe image with high-resolution gallery images registered in a database. Existing methods mainly leverage prior information from high-resolution images by either reconstructing facial details with super-resolution techniques or learning a unified feature space. To address this challenge, this paper proposes a new approach that enforces the network to focus on the discriminative information stored in the low-frequency components of a low-resolution image. A cross-resolution knowledge distillation paradigm is first employed as the learning framework. Then, an identity-preserving network, WaveResNet, and a wavelet similarity loss are designed to capture low-frequency details and boost performance. Finally, an image degradation model is conceived to simulate more realistic low-resolution training data. Consequently, extensive experimental results show that the proposed method consistently outperforms the baseline model and other state-of-the-art methods across a variety of image resolutions.
Algorithmic fairness has become an important machine learning problem, especially for mission-critical Web applications. This work presents a self-supervised model, called DualFair, that can debias sensitive attributes like gender and race from learned representations. Unlike existing models that target a single type of fairness, our model jointly optimizes for two fairness criteria - group fairness and counterfactual fairness - and hence makes fairer predictions at both the group and individual levels. Our model uses contrastive loss to generate embeddings that are indistinguishable for each protected group, while forcing the embeddings of counterfactual pairs to be similar. It then uses a self-knowledge distillation method to maintain the quality of representation for the downstream tasks. Extensive analysis over multiple datasets confirms the model's validity and further shows the synergy of jointly addressing two fairness criteria, suggesting the model's potential value in fair intelligent Web applications.
Online knowledge distillation (KD) has received increasing attention in recent years. However, while most existing online KD methods focus on developing complicated model structures and training strategies to improve the distillation of high-level knowledge like probability distribution, the effects of the multi-level knowledge in the online KD are greatly overlooked, especially the low-level knowledge. Thus, to provide a novel viewpoint to online KD, we propose MetaMixer, a regularization strategy that can strengthen the distillation by combining the low-level knowledge that impacts the localization capability of the networks, and high-level knowledge that focuses on the whole image. Experiments under different conditions show that MetaMixer can achieve significant performance gains over state-of-the-art methods.
This paper tackles the problem of semi-supervised video object segmentation on resource-constrained devices, such as mobile phones. We formulate this problem as a distillation task, whereby we demonstrate that small space-time-memory networks with finite memory can achieve competitive results with state of the art, but at a fraction of the computational cost (32 milliseconds per frame on a Samsung Galaxy S22). Specifically, we provide a theoretically grounded framework that unifies knowledge distillation with supervised contrastive representation learning. These models are able to jointly benefit from both pixel-wise contrastive learning and distillation from a pre-trained teacher. We validate this loss by achieving competitive J&F to state of the art on both the standard DAVIS and YouTube benchmarks, despite running up to 5x faster, and with 32x fewer parameters.
Data-Free Knowledge Distillation (DFKD) has recently attracted growing attention in the academic community, especially with major breakthroughs in computer vision. Despite promising results, the technique has not been well applied to audio and signal processing. Due to the variable duration of audio signals, it has its own unique way of modeling. In this work, we propose feature-rich audio model inversion (FRAMI), a data-free knowledge distillation framework for general sound classification tasks. It first generates high-quality and feature-rich Mel-spectrograms through a feature-invariant contrastive loss. Then, the hidden states before and after the statistics pooling layer are reused when knowledge distillation is performed on these feature-rich samples. Experimental results on the Urbansound8k, ESC-50, and audioMNIST datasets demonstrate that FRAMI can generate feature-rich samples. Meanwhile, the accuracy of the student model is further improved by reusing the hidden state and significantly outperforms the baseline method.
Many radar signal processing methodologies are being developed for critical road safety perception tasks. Unfortunately, these signal processing algorithms are often poorly suited to run on embedded hardware accelerators used in automobiles. Conversely, end-to-end machine learning (ML) approaches better exploit the performance gains brought by specialized accelerators. In this paper, we propose a teacher-student knowledge distillation approach for low-level radar perception tasks. We utilize a hybrid model for stationary object detection as a teacher to train an end-to-end ML student model. The student can efficiently harness embedded compute for real-time deployment. We demonstrate that the proposed student model runs at speeds 100x faster than the teacher model.
Since large number of high-quality remote sensing images are readily accessible, exploiting the corpus of images with less manual annotation draws increasing attention. Self-supervised models acquire general feature representations by formulating a pretext task that generates pseudo-labels for massive unlabeled data to provide supervision for training. While prior studies have explored multiple self-supervised learning techniques in remote sensing domain, pretext tasks based on local-global view alignment remain underexplored, despite achieving state-of-the-art results on natural imagery. Inspired by DINO, which employs an effective representation learning structure with knowledge distillation based on global-local view alignment, we formulate two pretext tasks for self-supervised learning on remote sensing imagery (SSLRS). Using these tasks, we explore the effectiveness of positive temporal contrast as well as multi-sized views on SSLRS. We extend DINO and propose DINO-MC which uses local views of various sized crops instead of a single fixed size in order to alleviate the limited variation in object size observed in remote sensing imagery. Our experiments demonstrate that even when pre-trained on only 10% of the dataset, DINO-MC performs on par or better than existing state-of-the-art SSLRS methods on multiple remote sensing tasks, while using less computational resources. All codes, models, and results are released at https://github.com/WennyXY/DINO-MC.
In many practical scenarios -- like hyperparameter search or continual retraining with new data -- related training runs are performed many times in sequence. Current practice is to train each of these models independently from scratch. We study the problem of exploiting the computation invested in previous runs to reduce the cost of future runs using knowledge distillation (KD). We find that augmenting future runs with KD from previous runs dramatically reduces the time necessary to train these models, even taking into account the overhead of KD. We improve on these results with two strategies that reduce the overhead of KD by 80-90% with minimal effect on accuracy and vast pareto-improvements in overall cost. We conclude that KD is a promising avenue for reducing the cost of the expensive preparatory work that precedes training final models in practice.
Training deep models for semantic scene completion (SSC) is challenging due to the sparse and incomplete input, a large quantity of objects of diverse scales as well as the inherent label noise for moving objects. To address the above-mentioned problems, we propose the following three solutions: 1) Redesigning the completion sub-network. We design a novel completion sub-network, which consists of several Multi-Path Blocks (MPBs) to aggregate multi-scale features and is free from the lossy downsampling operations. 2) Distilling rich knowledge from the multi-frame model. We design a novel knowledge distillation objective, dubbed Dense-to-Sparse Knowledge Distillation (DSKD). It transfers the dense, relation-based semantic knowledge from the multi-frame teacher to the single-frame student, significantly improving the representation learning of the single-frame model. 3) Completion label rectification. We propose a simple yet effective label rectification strategy, which uses off-the-shelf panoptic segmentation labels to remove the traces of dynamic objects in completion labels, greatly improving the performance of deep models especially for those moving objects. Extensive experiments are conducted in two public SSC benchmarks, i.e., SemanticKITTI and SemanticPOSS. Our SCPNet ranks 1st on SemanticKITTI semantic scene completion challenge and surpasses the competitive S3CNet by 7.2 mIoU. SCPNet also outperforms previous completion algorithms on the SemanticPOSS dataset. Besides, our method also achieves competitive results on SemanticKITTI semantic segmentation tasks, showing that knowledge learned in the scene completion is beneficial to the segmentation task.
The goal of continuous sign language recognition(CSLR) research is to apply CSLR models as a communication tool in real life, and the real-time requirement of the models is important. In this paper, we address the model real-time problem through cross-resolution knowledge distillation. In our study, we found that keeping the frame-level feature scales consistent between the output of the student network and the teacher network is better than recovering the frame-level feature sizes for feature distillation. Based on this finding, we propose a new frame-level feature extractor that keeps the output frame-level features at the same scale as the output of by the teacher network. We further combined with the TSCM+2D hybrid convolution proposed in our previous study to form a new lightweight end-to-end CSLR network-Low resolution input net(LRINet). It is then used to combine cross-resolution knowledge distillation and traditional knowledge distillation methods to form a CSLR model based on cross-resolution knowledge distillation (CRKD). The CRKD uses high-resolution frames as input to the teacher network for training, locks the weights after training, and then uses low-resolution frames as input to the student network LRINet to perform knowledge distillation on frame-level features and classification features respectively. Experiments on two large-scale continuous sign language datasets have proved the effectiveness of CRKD. Compared with the model with high-resolution data as input, the calculation amount, parameter amount and inference time of the model have been significantly reduced under the same experimental conditions, while ensuring the accuracy of the model, and has achieved very competitive results in comparison with other advanced methods.
In this paper, to deal with the heterogeneity in federated learning (FL) systems, a knowledge distillation (KD) driven training framework for FL is proposed, where each user can select its neural network model on demand and distill knowledge from a big teacher model using its own private dataset. To overcome the challenge of train the big teacher model in resource limited user devices, the digital twin (DT) is exploit in the way that the teacher model can be trained at DT located in the server with enough computing resources. Then, during model distillation, each user can update the parameters of its model at either the physical entity or the digital agent. The joint problem of model selection and training offloading and resource allocation for users is formulated as a mixed integer programming (MIP) problem. To solve the problem, Q-learning and optimization are jointly used, where Q-learning selects models for users and determines whether to train locally or on the server, and optimization is used to allocate resources for users based on the output of Q-learning. Simulation results show the proposed DT-assisted KD framework and joint optimization method can significantly improve the average accuracy of users while reducing the total delay.
The use of multi-camera views simultaneously has been shown to improve the generalization capabilities and performance of visual policies. However, the hardware cost and design constraints in real-world scenarios can potentially make it challenging to use multiple cameras. In this study, we present a novel approach to enhance the generalization performance of vision-based Reinforcement Learning (RL) algorithms for robotic manipulation tasks. Our proposed method involves utilizing a technique known as knowledge distillation, in which a pre-trained ``teacher'' policy trained with multiple camera viewpoints guides a ``student'' policy in learning from a single camera viewpoint. To enhance the student policy's robustness against camera location perturbations, it is trained using data augmentation and extreme viewpoint changes. As a result, the student policy learns robust visual features that allow it to locate the object of interest accurately and consistently, regardless of the camera viewpoint. The efficacy and efficiency of the proposed method were evaluated both in simulation and real-world environments. The results demonstrate that the single-view visual student policy can successfully learn to grasp and lift a challenging object, which was not possible with a single-view policy alone. Furthermore, the student policy demonstrates zero-shot transfer capability, where it can successfully grasp and lift objects in real-world scenarios for unseen visual configurations.
Social bot detection is of paramount importance to the resilience and security of online social platforms. The state-of-the-art detection models are siloed and have largely overlooked a variety of data characteristics from multiple cross-lingual platforms. Meanwhile, the heterogeneity of data distribution and model architecture makes it intricate to devise an efficient cross-platform and cross-model detection framework. In this paper, we propose FedACK, a new federated adversarial contrastive knowledge distillation framework for social bot detection. We devise a GAN-based federated knowledge distillation mechanism for efficiently transferring knowledge of data distribution among clients. In particular, a global generator is used to extract the knowledge of global data distribution and distill it into each client's local model. We leverage local discriminator to enable customized model design and use local generator for data enhancement with hard-to-decide samples. Local training is conducted as multi-stage adversarial and contrastive learning to enable consistent feature spaces among clients and to constrain the optimization direction of local models, reducing the divergences between local and global models. Experiments demonstrate that FedACK outperforms the state-of-the-art approaches in terms of accuracy, communication efficiency, and feature space consistency.
This work studies knowledge distillation (KD) and addresses its constraints for recurrent neural network transducer (RNN-T) models. In hard distillation, a teacher model transcribes large amounts of unlabelled speech to train a student model. Soft distillation is another popular KD method that distills the output logits of the teacher model. Due to the nature of RNN-T alignments, applying soft distillation between RNN-T architectures having different posterior distributions is challenging. In addition, bad teachers having high word-error-rate (WER) reduce the efficacy of KD. We investigate how to effectively distill knowledge from variable quality ASR teachers, which has not been studied before to the best of our knowledge. We show that a sequence-level KD, full-sum distillation, outperforms other distillation methods for RNN-T models, especially for bad teachers. We also propose a variant of full-sum distillation that distills the sequence discriminative knowledge of the teacher leading to further improvement in WER. We conduct experiments on public datasets namely SpeechStew and LibriSpeech, and on in-house production data.
Despite the success of deep learning models on instance segmentation, current methods still suffer from catastrophic forgetting in continual learning scenarios. In this paper, our contributions for continual instance segmentation are threefold. First, we propose the Y-knowledge distillation (Y-KD), a technique that shares a common feature extractor between the teacher and student networks. As the teacher is also updated with new data in Y-KD, the increased plasticity results in new modules that are specialized on new classes. Second, our Y-KD approach is supported by a dynamic architecture method that trains task-specific modules with a unique instance segmentation head, thereby significantly reducing forgetting. Third, we complete our approach by leveraging checkpoint averaging as a simple method to manually balance the trade-off between performance on the various sets of classes, thus increasing control over the model's behavior without any additional cost. These contributions are united in our model that we name the Dynamic Y-KD network. We perform extensive experiments on several single-step and multi-steps incremental learning scenarios, and we show that our approach outperforms previous methods both on past and new classes. For instance, compared to recent work, our method obtains +2.1% mAP on old classes in 15-1, +7.6% mAP on new classes in 19-1 and reaches 91.5% of the mAP obtained by joint-training on all classes in 15-5.