Topic modeling is a dominant method for exploring document collections on the web and in digital libraries. Recent approaches to topic modeling use pretrained contextualized language models and variational autoencoders. However, large neural topic models have a considerable memory footprint. In this paper, we propose a knowledge distillation framework to compress a contextualized topic model without loss in topic quality. In particular, the proposed distillation objective is to minimize the cross-entropy of the soft labels produced by the teacher and the student models, as well as to minimize the squared 2-Wasserstein distance between the latent distributions learned by the two models. Experiments on two publicly available datasets show that the student trained with knowledge distillation achieves topic coherence much higher than that of the original student model, and even surpasses the teacher while containing far fewer parameters than the teacher's. The distilled model also outperforms several other competitive topic models on topic coherence.
The size and the computational load of fine-tuning large-scale pre-trained neural network are becoming two major obstacles in adopting machine learning in many applications. Continual learning (CL) can serve as a remedy through enabling knowledge-transfer across sequentially arriving tasks which relaxes the need to fine-tune all network weights from scratch. However, existing CL algorithms primarily consider learning unimodal vision-only or language-only tasks. We develop a transformer-based CL architecture for learning bimodal vision-and-language tasks based on increasing the number of the learnable parameters dynamically and using knowledge distillation. The new additional parameters are used to specialize the network for each task. Our approach enables sharing information between the tasks while addressing the challenge of catastrophic forgetting. Our approach is scalable learning to a large number of tasks because it requires little memory and time overhead. Our model reaches state-of-the-art performance on challenging vision-and-language tasks.
Light Field Networks, the re-formulations of radiance fields to oriented rays, are magnitudes faster than their coordinate network counterparts, and provide higher fidelity with respect to representing 3D structures from 2D observations. They would be well suited for generic scene representation and manipulation, but suffer from one problem: they are limited to holistic and static scenes. In this paper, we propose the Dynamic Light Field Network (DyLiN) method that can handle non-rigid deformations, including topological changes. We learn a deformation field from input rays to canonical rays, and lift them into a higher dimensional space to handle discontinuities. We further introduce CoDyLiN, which augments DyLiN with controllable attribute inputs. We train both models via knowledge distillation from pretrained dynamic radiance fields. We evaluated DyLiN using both synthetic and real world datasets that include various non-rigid deformations. DyLiN qualitatively outperformed and quantitatively matched state-of-the-art methods in terms of visual fidelity, while being 25 - 71x computationally faster. We also tested CoDyLiN on attribute annotated data and it surpassed its teacher model. Project page: https://dylin2023.github.io .
Most existing online knowledge distillation(OKD) techniques typically require sophisticated modules to produce diverse knowledge for improving students' generalization ability. In this paper, we strive to fully utilize multi-model settings instead of well-designed modules to achieve a distillation effect with excellent generalization performance. Generally, model generalization can be reflected in the flatness of the loss landscape. Since averaging parameters of multiple models can find flatter minima, we are inspired to extend the process to the sampled convex combinations of multi-student models in OKD. Specifically, by linearly weighting students' parameters in each training batch, we construct a Hybrid-Weight Model(HWM) to represent the parameters surrounding involved students. The supervision loss of HWM can estimate the landscape's curvature of the whole region around students to measure the generalization explicitly. Hence we integrate HWM's loss into students' training and propose a novel OKD framework via parameter hybridization(OKDPH) to promote flatter minima and obtain robust solutions. Considering the redundancy of parameters could lead to the collapse of HWM, we further introduce a fusion operation to keep the high similarity of students. Compared to the state-of-the-art(SOTA) OKD methods and SOTA methods of seeking flat minima, our OKDPH achieves higher performance with fewer parameters, benefiting OKD with lightweight and robust characteristics. Our code is publicly available at https://github.com/tianlizhang/OKDPH.
Recently, Transformer-based methods have been utilized to improve the performance of human action recognition. However, most of these studies assume that multi-view data is complete, which may not always be the case in real-world scenarios. Therefore, this paper presents a novel Multi-view Knowledge Distillation Transformer (MKDT) framework that consists of a teacher network and a student network. This framework aims to handle incomplete human action problems in real-world applications. Specifically, the multi-view knowledge distillation transformer uses a hierarchical vision transformer with shifted windows to capture more spatial-temporal information. Experimental results demonstrate that our framework outperforms the CNN-based method on three public datasets.
Federated Learning has gained popularity among medical institutions since it enables collaborative training between clients (e.g., hospitals) without aggregating data. However, due to the high cost associated with creating annotations, especially for large 3D image datasets, clinical institutions do not have enough supervised data for training locally. Thus, the performance of the collaborative model is subpar under limited supervision. On the other hand, large institutions have the resources to compile data repositories with high-resolution images and labels. Therefore, individual clients can utilize the knowledge acquired in the public data repositories to mitigate the shortage of private annotated images. In this paper, we propose a federated few-shot learning method with dual knowledge distillation. This method allows joint training with limited annotations across clients without jeopardizing privacy. The supervised learning of the proposed method extracts features from limited labeled data in each client, while the unsupervised data is used to distill both feature and response-based knowledge from a national data repository to further improve the accuracy of the collaborative model and reduce the communication cost. Extensive evaluations are conducted on 3D magnetic resonance knee images from a private clinical dataset. Our proposed method shows superior performance and less training time than other semi-supervised federated learning methods. Codes and additional visualization results are available at https://github.com/hexiaoxiao-cs/fedml-knee.
In the field of 3D object detection for autonomous driving, the sensor portfolio including multi-modality and single-modality is diverse and complex. Since the multi-modal methods have system complexity while the accuracy of single-modal ones is relatively low, how to make a tradeoff between them is difficult. In this work, we propose a universal cross-modality knowledge distillation framework (UniDistill) to improve the performance of single-modality detectors. Specifically, during training, UniDistill projects the features of both the teacher and the student detector into Bird's-Eye-View (BEV), which is a friendly representation for different modalities. Then, three distillation losses are calculated to sparsely align the foreground features, helping the student learn from the teacher without introducing additional cost during inference. Taking advantage of the similar detection paradigm of different detectors in BEV, UniDistill easily supports LiDAR-to-camera, camera-to-LiDAR, fusion-to-LiDAR and fusion-to-camera distillation paths. Furthermore, the three distillation losses can filter the effect of misaligned background information and balance between objects of different sizes, improving the distillation effectiveness. Extensive experiments on nuScenes demonstrate that UniDistill effectively improves the mAP and NDS of student detectors by 2.0%~3.2%.
This paper investigates cross-lingual temporal knowledge graph reasoning problem, which aims to facilitate reasoning on Temporal Knowledge Graphs (TKGs) in low-resource languages by transfering knowledge from TKGs in high-resource ones. The cross-lingual distillation ability across TKGs becomes increasingly crucial, in light of the unsatisfying performance of existing reasoning methods on those severely incomplete TKGs, especially in low-resource languages. However, it poses tremendous challenges in two aspects. First, the cross-lingual alignments, which serve as bridges for knowledge transfer, are usually too scarce to transfer sufficient knowledge between two TKGs. Second, temporal knowledge discrepancy of the aligned entities, especially when alignments are unreliable, can mislead the knowledge distillation process. We correspondingly propose a mutually-paced knowledge distillation model MP-KD, where a teacher network trained on a source TKG can guide the training of a student network on target TKGs with an alignment module. Concretely, to deal with the scarcity issue, MP-KD generates pseudo alignments between TKGs based on the temporal information extracted by our representation module. To maximize the efficacy of knowledge transfer and control the noise caused by the temporal knowledge discrepancy, we enhance MP-KD with a temporal cross-lingual attention mechanism to dynamically estimate the alignment strength. The two procedures are mutually paced along with model training. Extensive experiments on twelve cross-lingual TKG transfer tasks in the EventKG benchmark demonstrate the effectiveness of the proposed MP-KD method.
Catastrophic forgetting has been a major challenge in continual learning, where the model needs to learn new tasks with limited or no access to data from previously seen tasks. To tackle this challenge, methods based on knowledge distillation in feature space have been proposed and shown to reduce forgetting. However, most feature distillation methods directly constrain the new features to match the old ones, overlooking the need for plasticity. To achieve a better stability-plasticity trade-off, we propose Backward Feature Projection (BFP), a method for continual learning that allows the new features to change up to a learnable linear transformation of the old features. BFP preserves the linear separability of the old classes while allowing the emergence of new feature directions to accommodate new classes. BFP can be integrated with existing experience replay methods and boost performance by a significant margin. We also demonstrate that BFP helps learn a better representation space, in which linear separability is well preserved during continual learning and linear probing achieves high classification accuracy. The code can be found at https://github.com/rvl-lab-utoronto/BFP
Multi-frame methods improve monocular depth estimation over single-frame approaches by aggregating spatial-temporal information via feature matching. However, the spatial-temporal feature leads to accuracy degradation in dynamic scenes. To enhance the performance, recent methods tend to propose complex architectures for feature matching and dynamic scenes. In this paper, we show that a simple learning framework, together with designed feature augmentation, leads to superior performance. (1) A novel dynamic objects detecting method with geometry explainability is proposed. The detected dynamic objects are excluded during training, which guarantees the static environment assumption and relieves the accuracy degradation problem of the multi-frame depth estimation. (2) Multi-scale feature fusion is proposed for feature matching in the multi-frame depth network, which improves feature matching, especially between frames with large camera motion. (3) The robust knowledge distillation with a robust teacher network and reliability guarantee is proposed, which improves the multi-frame depth estimation without computation complexity increase during the test. The experiments show that our proposed methods achieve great performance improvement on the multi-frame depth estimation.
Manufacturing wafers is an intricate task involving thousands of steps. Defect Pattern Recognition (DPR) of wafer maps is crucial for determining the root cause of production defects, which may further provide insight for yield improvement in wafer foundry. During manufacturing, various defects may appear standalone in the wafer or may appear as different combinations. Identifying multiple defects in a wafer is generally harder compared to identifying a single defect. Recently, deep learning methods have gained significant traction in mixed-type DPR. However, the complexity of defects requires complex and large models making them very difficult to operate on low-memory embedded devices typically used in fabrication labs. Another common issue is the unavailability of labeled data to train complex networks. In this work, we propose an unsupervised training routine to distill the knowledge of complex pre-trained models to lightweight deployment-ready models. We empirically show that this type of training compresses the model without sacrificing accuracy despite being up to 10 times smaller than the teacher model. The compressed model also manages to outperform contemporary state-of-the-art models.
This paper advances the fine-grained sketch-based image retrieval (FG-SBIR) literature by putting forward a strong baseline that overshoots prior state-of-the-arts by ~11%. This is not via complicated design though, but by addressing two critical issues facing the community (i) the gold standard triplet loss does not enforce holistic latent space geometry, and (ii) there are never enough sketches to train a high accuracy model. For the former, we propose a simple modification to the standard triplet loss, that explicitly enforces separation amongst photos/sketch instances. For the latter, we put forward a novel knowledge distillation module can leverage photo data for model training. Both modules are then plugged into a novel plug-n-playable training paradigm that allows for more stable training. More specifically, for (i) we employ an intra-modal triplet loss amongst sketches to bring sketches of the same instance closer from others, and one more amongst photos to push away different photo instances while bringing closer a structurally augmented version of the same photo (offering a gain of ~4-6%). To tackle (ii), we first pre-train a teacher on the large set of unlabelled photos over the aforementioned intra-modal photo triplet loss. Then we distill the contextual similarity present amongst the instances in the teacher's embedding space to that in the student's embedding space, by matching the distribution over inter-feature distances of respective samples in both embedding spaces (delivering a further gain of ~4-5%). Apart from outperforming prior arts significantly, our model also yields satisfactory results on generalising to new classes. Project page: https://aneeshan95.github.io/Sketch_PVT/
Place recognition is an essential and challenging task in loop closing and global localization for robotics and autonomous driving applications. Benefiting from the recent advances in deep learning techniques, the performance of LiDAR place recognition (LPR) has been greatly improved. However, current deep learning-based methods suffer from two major problems: poor generalization ability and catastrophic forgetting. In this paper, we propose a continual contrastive learning method, named CCL, to tackle the catastrophic forgetting problem and generally improve the robustness of LPR approaches. Our CCL constructs a contrastive feature pool and utilizes contrastive loss to train more transferable representations of places. When transferred into new environments, our CCL continuously reviews the contrastive memory bank and applies a distribution-based knowledge distillation to maintain the retrieval ability of the past data while continually learning to recognize new places from the new data. We thoroughly evaluate our approach on Oxford, MulRan, and PNV datasets using three different LPR methods. The experimental results show that our CCL consistently improves the performance of different methods in different environments outperforming the state-of-the-art continual learning method. The implementation of our method has been released at https://github.com/cloudcjf/CCL.
Human multimodal emotion recognition (MER) aims to perceive human emotions via language, visual and acoustic modalities. Despite the impressive performance of previous MER approaches, the inherent multimodal heterogeneities still haunt and the contribution of different modalities varies significantly. In this work, we mitigate this issue by proposing a decoupled multimodal distillation (DMD) approach that facilitates flexible and adaptive crossmodal knowledge distillation, aiming to enhance the discriminative features of each modality. Specially, the representation of each modality is decoupled into two parts, i.e., modality-irrelevant/-exclusive spaces, in a self-regression manner. DMD utilizes a graph distillation unit (GD-Unit) for each decoupled part so that each GD can be performed in a more specialized and effective manner. A GD-Unit consists of a dynamic graph where each vertice represents a modality and each edge indicates a dynamic knowledge distillation. Such GD paradigm provides a flexible knowledge transfer manner where the distillation weights can be automatically learned, thus enabling diverse crossmodal knowledge transfer patterns. Experimental results show DMD consistently obtains superior performance than state-of-the-art MER methods. Visualization results show the graph edges in DMD exhibit meaningful distributional patterns w.r.t. the modality-irrelevant/-exclusive feature spaces. Codes are released at \url{https://github.com/mdswyz/DMD}.
The Area Under ROC curve (AUC) is widely used as an evaluation metric in various applications. Due to its insensitivity towards class distribution, directly optimizing AUC performs well on the class imbalance problem. However, existing AUC optimization methods are limited to regular data such as text, images, and video. AUC optimization on graph data, which is ubiquitous and important, is seldom studied. Different from regular data, AUC optimization on graphs suffers from not only the class imbalance but also topology imbalance. To solve the complicated imbalance problem, we propose a unified topology-aware AUC optimization (TOPOAUC) framework, which could simultaneously deal with the topology and class imbalance problem in graph learning. We develop a multi-class AUC optimization work to deal with the class imbalance problem. With respect to topology imbalance, we propose a T opology-A ware I mportance L earning mechanism (TAIL), which considers the topology of pairwise nodes and different contributions of topology information to pairwise node neighbors. Extensive experiments on three real-world datasets demonstrate the effectiveness of our proposed method.
Distilling high-accuracy Graph Neural Networks (GNNs) to low-latency multilayer perceptions (MLPs) on graph tasks has become a hot research topic. However, conventional MLP learning relies almost exclusively on graph nodes and fails to effectively capture the graph structural information. Previous methods address this issue by processing graph edges into extra inputs for MLPs, but such graph structures may be unavailable for various scenarios. To this end, we propose Prototype-Guided Knowledge Distillation (PGKD), which does not require graph edges (edge-free setting) yet learns structure-aware MLPs. Our insight is to distill graph structural information from GNNs. Specifically, we first employ the class prototypes to analyze the impact of graph structures on GNN teachers, and then design two losses to distill such information from GNNs to MLPs. Experimental results on popular graph benchmarks demonstrate the effectiveness and robustness of the proposed PGKD.
Knowledge distillation is a popular technique for transferring the knowledge from a large teacher model to a smaller student model by mimicking. However, distillation by directly aligning the feature maps between teacher and student may enforce overly strict constraints on the student thus degrade the performance of the student model. To alleviate the above feature misalignment issue, existing works mainly focus on spatially aligning the feature maps of the teacher and the student, with pixel-wise transformation. In this paper, we newly find that aligning the feature maps between teacher and student along the channel-wise dimension is also effective for addressing the feature misalignment issue. Specifically, we propose a learnable nonlinear channel-wise transformation to align the features of the student and the teacher model. Based on it, we further propose a simple and generic framework for feature distillation, with only one hyper-parameter to balance the distillation loss and the task specific loss. Extensive experimental results show that our method achieves significant performance improvements in various computer vision tasks including image classification (+3.28% top-1 accuracy for MobileNetV1 on ImageNet-1K), object detection (+3.9% bbox mAP for ResNet50-based Faster-RCNN on MS COCO), instance segmentation (+2.8% Mask mAP for ResNet50-based Mask-RCNN), and semantic segmentation (+4.66% mIoU for ResNet18-based PSPNet in semantic segmentation on Cityscapes), which demonstrates the effectiveness and the versatility of the proposed method. The code will be made publicly available.
Knowledge Distillation (KD) uses the teacher's prediction logits as soft labels to guide the student, while self-KD does not need a real teacher to require the soft labels. This work unifies the formulations of the two tasks by decomposing and reorganizing the generic KD loss into a Normalized KD (NKD) loss and customized soft labels for both target class (image's category) and non-target classes named Universal Self-Knowledge Distillation (USKD). We decompose the KD loss and find the non-target loss from it forces the student's non-target logits to match the teacher's, but the sum of the two non-target logits is different, preventing them from being identical. NKD normalizes the non-target logits to equalize their sum. It can be generally used for KD and self-KD to better use the soft labels for distillation loss. USKD generates customized soft labels for both target and non-target classes without a teacher. It smooths the target logit of the student as the soft target label and uses the rank of the intermediate feature to generate the soft non-target labels with Zipf's law. For KD with teachers, our NKD achieves state-of-the-art performance on CIFAR-100 and ImageNet datasets, boosting the ImageNet Top-1 accuracy of ResNet18 from 69.90% to 71.96% with a ResNet-34 teacher. For self-KD without teachers, USKD is the first self-KD method that can be effectively applied to both CNN and ViT models with negligible additional time and memory cost, resulting in new state-of-the-art results, such as 1.17% and 0.55% accuracy gains on ImageNet for MobileNet and DeiT-Tiny, respectively. Our codes are available at https://github.com/yzd-v/cls_KD.
Knowledge graph embedding (KGE) that maps entities and relations into vector representations is essential for downstream applications. Conventional KGE methods require high-dimensional representations to learn the complex structure of knowledge graph, but lead to oversized model parameters. Recent advances reduce parameters by low-dimensional entity representations, while developing techniques (e.g., knowledge distillation or reinvented representation forms) to compensate for reduced dimension. However, such operations introduce complicated computations and model designs that may not benefit large knowledge graphs. To seek a simple strategy to improve the parameter efficiency of conventional KGE models, we take inspiration from that deeper neural networks require exponentially fewer parameters to achieve expressiveness comparable to wider networks for compositional structures. We view all entity representations as a single-layer embedding network, and conventional KGE methods that adopt high-dimensional entity representations equal widening the embedding network to gain expressiveness. To achieve parameter efficiency, we instead propose a deeper embedding network for entity representations, i.e., a narrow entity embedding layer plus a multi-layer dimension lifting network (LiftNet). Experiments on three public datasets show that by integrating LiftNet, four conventional KGE methods with 16-dimensional representations achieve comparable link prediction accuracy as original models that adopt 512-dimensional representations, saving 68.4% to 96.9% parameters.
Recent open-vocabulary detection methods aim to detect novel objects by distilling knowledge from vision-language models (VLMs) trained on a vast amount of image-text pairs. To improve the effectiveness of these methods, researchers have utilized datasets with a large vocabulary that contains a large number of object classes, under the assumption that such data will enable models to extract comprehensive knowledge on the relationships between various objects and better generalize to unseen object classes. In this study, we argue that more fine-grained labels are necessary to extract richer knowledge about novel objects, including object attributes and relationships, in addition to their names. To address this challenge, we propose a simple and effective method named Pseudo Caption Labeling (PCL), which utilizes an image captioning model to generate captions that describe object instances from diverse perspectives. The resulting pseudo caption labels offer dense samples for knowledge distillation. On the LVIS benchmark, our best model trained on the de-duplicated VisualGenome dataset achieves an AP of 34.5 and an APr of 30.6, comparable to the state-of-the-art performance. PCL's simplicity and flexibility are other notable features, as it is a straightforward pre-processing technique that can be used with any image captioning model without imposing any restrictions on model architecture or training process.