Word-level handwritten optical character recognition (OCR) remains a challenge for morphologically rich languages like Bangla. The complexity arises from the existence of a large number of alphabets, the presence of several diacritic forms, and the appearance of complex conjuncts. The difficulty is exacerbated by the fact that some graphemes occur infrequently but remain indispensable, so addressing the class imbalance is required for satisfactory results. This paper addresses this issue by introducing two knowledge distillation methods: Leveraging Isolated Letter Accumulations By Ordering Teacher Insights (LILA-BOTI) and Super Teacher LILA-BOTI. In both cases, a Convolutional Recurrent Neural Network (CRNN) student model is trained with the dark knowledge gained from a printed isolated character recognition teacher model. We conducted inter-dataset testing on \emph{BN-HTRd} and \emph{BanglaWriting} as our evaluation protocol, thus setting up a challenging problem where the results would better reflect the performance on unseen data. Our evaluations achieved up to a 3.5% increase in the F1-Macro score for the minor classes and up to 4.5% increase in our overall word recognition rate when compared with the base model (No KD) and conventional KD.
Knowledge Distillation (KD) is a typical method for training a lightweight student model with the help of a well-trained teacher model. However, most KD methods require access to either the teacher's training data or model parameters, which is unrealistic. To tackle this problem, recent works study KD under data-free and black-box settings. Nevertheless, these works require a large number of queries to the teacher model, which incurs significant monetary and computational costs. To address these problems, we propose a novel method called \emph{query-effIcient Data-free lEarning from blAck-box modeLs} (IDEAL), which aims to query-efficiently learn from black-box model APIs to train a good student without any real data. In detail, IDEAL trains the student model in two stages: data generation and model distillation. Note that IDEAL does not require any query in the data generation stage and queries the teacher only once for each sample in the distillation stage. Extensive experiments on various real-world datasets show the effectiveness of the proposed IDEAL. For instance, IDEAL can improve the performance of the best baseline method DFME by 5.83% on CIFAR10 dataset with only 0.02x the query budget of DFME.
Multi-Label Image Classification (MLIC) approaches usually exploit label correlations to achieve good performance. However, emphasizing correlation like co-occurrence may overlook discriminative features of the target itself and lead to model overfitting, thus undermining the performance. In this study, we propose a generic framework named Parallel Self-Distillation (PSD) for boosting MLIC models. PSD decomposes the original MLIC task into several simpler MLIC sub-tasks via two elaborated complementary task decomposition strategies named Co-occurrence Graph Partition (CGP) and Dis-occurrence Graph Partition (DGP). Then, the MLIC models of fewer categories are trained with these sub-tasks in parallel for respectively learning the joint patterns and the category-specific patterns of labels. Finally, knowledge distillation is leveraged to learn a compact global ensemble of full categories with these learned patterns for reconciling the label correlation exploitation and model overfitting. Extensive results on MS-COCO and NUS-WIDE datasets demonstrate that our framework can be easily plugged into many MLIC approaches and improve performances of recent state-of-the-art approaches. The explainable visual study also further validates that our method is able to learn both the category-specific and co-occurring features. The source code is released at https://github.com/Robbie-Xu/CPSD.
Recent work explored the potential of large-scale Transformer-based pre-trained models, especially Pre-trained Language Models (PLMs) in natural language processing. This raises many concerns from various perspectives, e.g., financial costs and carbon emissions. Compressing PLMs like BERT with negligible performance loss for faster inference and cheaper deployment has attracted much attention. In this work, we aim to explore larger compression ratios for PLMs, among which tensor decomposition is a potential but under-investigated one. Two decomposition and reconstruction protocols are further proposed to improve the effectiveness and efficiency during compression. Our compressed BERT with ${1}/{7}$ parameters in Transformer layers performs on-par with, sometimes slightly better than the original BERT in GLUE benchmark. A tiny version achieves $96.7\%$ performance of BERT-base with $ {1}/{48} $ encoder parameters (i.e., less than 2M parameters excluding the embedding layer) and $2.7 \times$ faster on inference. To show that the proposed method is orthogonal to existing compression methods like knowledge distillation, we also explore the benefit of the proposed method on a distilled BERT.
In this paper, we introduce InDistill, a method that serves as a warmup stage for enhancing Knowledge Distillation (KD) effectiveness. InDistill focuses on transferring critical information flow paths from a heavyweight teacher to a lightweight student. This is achieved via a training scheme based on curriculum learning that considers the distillation difficulty of each layer and the critical learning periods when the information flow paths are established. This procedure can lead to a student model that is better prepared to learn from the teacher. To ensure the applicability of InDistill across a wide range of teacher-student pairs, we also incorporate a pruning operation when there is a discrepancy in the width of the teacher and student layers. This pruning operation reduces the width of the teacher's intermediate layers to match those of the student, allowing direct distillation without the need for an encoding stage. The proposed method is extensively evaluated using various pairs of teacher-student architectures on CIFAR-10, CIFAR-100, and ImageNet datasets demonstrating that preserving the information flow paths consistently increases the performance of the baseline KD approaches on both classification and retrieval settings. The code is available at https://github.com/gsarridis/InDistill.
Neural retrievers based on pre-trained language models (PLMs), such as dual-encoders, have achieved promising performance on the task of open-domain question answering (QA). Their effectiveness can further reach new state-of-the-arts by incorporating cross-architecture knowledge distillation. However, most of the existing studies just directly apply conventional distillation methods. They fail to consider the particular situation where the teacher and student have different structures. In this paper, we propose a novel distillation method that significantly advances cross-architecture distillation for dual-encoders. Our method 1) introduces a self on-the-fly distillation method that can effectively distill late interaction (i.e., ColBERT) to vanilla dual-encoder, and 2) incorporates a cascade distillation process to further improve the performance with a cross-encoder teacher. Extensive experiments are conducted to validate that our proposed solution outperforms strong baselines and establish a new state-of-the-art on open-domain QA benchmarks.
Considering uncertainty estimation of modern neural networks (NNs) is one of the most important steps towards deploying machine learning systems to meaningful real-world applications such as in medicine, finance or autonomous systems. At the moment, ensembles of different NNs constitute the state-of-the-art in both accuracy and uncertainty estimation in different tasks. However, ensembles of NNs are unpractical under real-world constraints, since their computation and memory consumption scale linearly with the size of the ensemble, which increase their latency and deployment cost. In this work, we examine a simple regularisation approach for distribution-free knowledge distillation of ensemble of machine learning models into a single NN. The aim of the regularisation is to preserve the diversity, accuracy and uncertainty estimation characteristics of the original ensemble without any intricacies, such as fine-tuning. We demonstrate the generality of the approach on combinations of toy data, SVHN/CIFAR-10, simple to complex NN architectures and different tasks.
Transformer models have been developed in molecular science with excellent performance in applications including quantitative structure-activity relationship (QSAR) and virtual screening (VS). Compared with other types of models, however, they are large, which results in a high hardware requirement to abridge time for both training and inference processes. In this work, cross-layer parameter sharing (CLPS), and knowledge distillation (KD) are used to reduce the sizes of transformers in molecular science. Both methods not only have competitive QSAR predictive performance as compared to the original BERT model, but also are more parameter efficient. Furthermore, by integrating CLPS and KD into a two-state chemical network, we introduce a new deep lite chemical transformer model, DeLiCaTe. DeLiCaTe captures general-domains as well as task-specific knowledge, which lead to a 4x faster rate of both training and inference due to a 10- and 3-times reduction of the number of parameters and layers, respectively. Meanwhile, it achieves comparable performance in QSAR and VS modeling. Moreover, we anticipate that the model compression strategy provides a pathway to the creation of effective generative transformer models for organic drug and material design.
Data-free knowledge distillation (DFKD) conducts knowledge distillation via eliminating the dependence of original training data, and has recently achieved impressive results in accelerating pre-trained language models. At the heart of DFKD is to reconstruct a synthetic dataset by inverting the parameters of the uncompressed model. Prior DFKD approaches, however, have largely relied on hand-crafted priors of the target data distribution for the reconstruction, which can be inevitably biased and often incompetent to capture the intrinsic distributions. To address this problem, we propose a prompt-based method, termed as PromptDFD, that allows us to take advantage of learned language priors, which effectively harmonizes the synthetic sentences to be semantically and grammatically correct. Specifically, PromptDFD leverages a pre-trained generative model to provide language priors and introduces a reinforced topic prompter to control data synthesis, making the generated samples thematically relevant and semantically plausible, and thus friendly to downstream tasks. As shown in our experiments, the proposed method substantially improves the synthesis quality and achieves considerable improvements on distillation performance. In some cases, PromptDFD even gives rise to results on par with those from the data-driven knowledge distillation with access to the original training data.
Non-autoregressive Transformers (NATs) significantly reduce the decoding latency by generating all tokens in parallel. However, such independent predictions prevent NATs from capturing the dependencies between the tokens for generating multiple possible translations. In this paper, we propose Directed Acyclic Transfomer (DA-Transformer), which represents the hidden states in a Directed Acyclic Graph (DAG), where each path of the DAG corresponds to a specific translation. The whole DAG simultaneously captures multiple translations and facilitates fast predictions in a non-autoregressive fashion. Experiments on the raw training data of WMT benchmark show that DA-Transformer substantially outperforms previous NATs by about 3 BLEU on average, which is the first NAT model that achieves competitive results with autoregressive Transformers without relying on knowledge distillation.
Machine learning models are prone to overfitting their training (source) domains, which is commonly believed to be the reason why they falter in novel target domains. Here we examine the contrasting view that multi-source domain generalization (DG) is first and foremost a problem of mitigating source domain underfitting: models not adequately learning the signal already present in their multi-domain training data. Experiments on a reading comprehension DG benchmark show that as a model learns its source domains better -- using familiar methods such as knowledge distillation (KD) from a bigger model -- its zero-shot out-of-domain utility improves at an even faster pace. Improved source domain learning also demonstrates superior out-of-domain generalization over three popular existing DG approaches that aim to limit overfitting. Our implementation of KD-based domain generalization is available via PrimeQA at: https://ibm.biz/domain-generalization-with-kd.
Existing knowledge distillation methods mostly focus on distillation of teacher's prediction and intermediate activation. However, the structured representation, which arguably is one of the most critical ingredients of deep models, is largely overlooked. In this work, we propose a novel {\em \modelname{}} ({\bf\em \shortname{})} method dedicated for distilling representational knowledge semantically from a pretrained teacher to a target student. The key idea is that we leverage the teacher's classifier as a semantic critic for evaluating the representations of both teacher and student and distilling the semantic knowledge with high-order structured information over all feature dimensions. This is accomplished by introducing a notion of cross-network logit computed through passing student's representation into teacher's classifier. Further, considering the set of seen classes as a basis for the semantic space in a combinatorial perspective, we scale \shortname{} to unseen classes for enabling effective exploitation of largely available, arbitrary unlabeled training data. At the problem level, this establishes an interesting connection between knowledge distillation with open-set semi-supervised learning (SSL). Extensive experiments show that our \shortname{} outperforms significantly previous state-of-the-art knowledge distillation methods on both coarse object classification and fine face recognition tasks, as well as less studied yet practically crucial binary network distillation. Under more realistic open-set SSL settings we introduce, we reveal that knowledge distillation is generally more effective than existing Out-Of-Distribution (OOD) sample detection, and our proposed \shortname{} is superior over both previous distillation and SSL competitors. The source code is available at \url{https://github.com/jingyang2017/SRD\_ossl}.
Despite the recent success of deep learning architectures, person re-identification (ReID) remains a challenging problem in real-word applications. Several unsupervised single-target domain adaptation (STDA) methods have recently been proposed to limit the decline in ReID accuracy caused by the domain shift that typically occurs between source and target video data. Given the multimodal nature of person ReID data (due to variations across camera viewpoints and capture conditions), training a common CNN backbone to address domain shifts across multiple target domains, can provide an efficient solution for real-time ReID applications. Although multi-target domain adaptation (MTDA) has not been widely addressed in the ReID literature, a straightforward approach consists in blending different target datasets, and performing STDA on the mixture to train a common CNN. However, this approach may lead to poor generalization, especially when blending a growing number of distinct target domains to train a smaller CNN. To alleviate this problem, we introduce a new MTDA method based on knowledge distillation (KD-ReID) that is suitable for real-time person ReID applications. Our method adapts a common lightweight student backbone CNN over the target domains by alternatively distilling from multiple specialized teacher CNNs, each one adapted on data from a specific target domain. Extensive experiments conducted on several challenging person ReID datasets indicate that our approach outperforms state-of-art methods for MTDA, including blending methods, particularly when training a compact CNN backbone like OSNet. Results suggest that our flexible MTDA approach can be employed to design cost-effective ReID systems for real-time video surveillance applications.
As an important and challenging problem, few-shot image generation aims at generating realistic images through training a GAN model given few samples. A typical solution for few-shot generation is to transfer a well-trained GAN model from a data-rich source domain to the data-deficient target domain. In this paper, we propose a novel self-supervised transfer scheme termed D3T-GAN, addressing the cross-domain GANs transfer in few-shot image generation. Specifically, we design two individual strategies to transfer knowledge between generators and discriminators, respectively. To transfer knowledge between generators, we conduct a data-dependent transformation, which projects and reconstructs the target samples into the source generator space. Then, we perform knowledge transfer from transformed samples to generated samples. To transfer knowledge between discriminators, we design a multi-level discriminant knowledge distillation from the source discriminator to the target discriminator on both the real and fake samples. Extensive experiments show that our method improve the quality of generated images and achieves the state-of-the-art FID scores on commonly used datasets.
We present TIPSy-GAN, a new approach to improve the accuracy and stability in unsupervised adversarial 2D to 3D human pose estimation. In our work we demonstrate that the human kinematic skeleton should not be assumed as a single spatially codependent structure; in fact, we posit when a full 2D pose is provided during training, there is an inherent bias learned where the 3D coordinate of a keypoint is spatially codependent on the 2D coordinates of all other keypoints. To investigate our hypothesis we follow previous adversarial approaches but train two generators on spatially independent parts of the kinematic skeleton, the torso and the legs. We find that improving the self-consistency cycle is key to lowering the evaluation error and therefore introduce new consistency constraints during training. A TIPSy model is produced via knowledge distillation from these generators which can predict the 3D ordinates for the entire 2D pose with improved results. Furthermore, we address an unanswered question in prior work of how long to train in a truly unsupervised scenario. We show that for two independent generators training adversarially has improved stability than that of a solo generator which collapses. TIPSy decreases the average error by 17\% when compared to that of a baseline solo generator on the Human3.6M dataset. TIPSy improves upon other unsupervised approaches while also performing strongly against supervised and weakly-supervised approaches during evaluation on both the Human3.6M and MPI-INF-3DHP datasets.
Graph neural networks (GNNs) have been widely used in modeling graph structured data, owing to its impressive performance in a wide range of practical applications. Recently, knowledge distillation (KD) for GNNs has enabled remarkable progress in graph model compression and knowledge transfer. However, most of the existing KD methods require a large volume of real data, which are not readily available in practice, and may preclude their applicability in scenarios where the teacher model is trained on rare or hard to acquire datasets. To address this problem, we propose the first end-to-end framework for data-free adversarial knowledge distillation on graph structured data (DFAD-GNN). To be specific, our DFAD-GNN employs a generative adversarial network, which mainly consists of three components: a pre-trained teacher model and a student model are regarded as two discriminators, and a generator is utilized for deriving training graphs to distill knowledge from the teacher model into the student model. Extensive experiments on various benchmark models and six representative datasets demonstrate that our DFAD-GNN significantly surpasses state-of-the-art data-free baselines in the graph classification task.
Knowledge distillation is widely adopted in semantic segmentation to reduce the computation cost.The previous knowledge distillation methods for semantic segmentation focus on pixel-wise feature alignment and intra-class feature variation distillation, neglecting to transfer the knowledge of the inter-class distance in the feature space, which is important for semantic segmentation. To address this issue, we propose an Inter-class Distance Distillation (IDD) method to transfer the inter-class distance in the feature space from the teacher network to the student network. Furthermore, semantic segmentation is a position-dependent task,thus we exploit a position information distillation module to help the student network encode more position information. Extensive experiments on three popular datasets: Cityscapes, Pascal VOC and ADE20K show that our method is helpful to improve the accuracy of semantic segmentation models and achieves the state-of-the-art performance. E.g. it boosts the benchmark model("PSPNet+ResNet18") by 7.50% in accuracy on the Cityscapes dataset.
Incremental few-shot object detection aims at detecting novel classes without forgetting knowledge of the base classes with only a few labeled training data from the novel classes. Most related prior works are on incremental object detection that rely on the availability of abundant training samples per novel class that substantially limits the scalability to real-world setting where novel data can be scarce. In this paper, we propose the Incremental-DETR that does incremental few-shot object detection via fine-tuning and self-supervised learning on the DETR object detector. To alleviate severe over-fitting with few novel class data, we first fine-tune the class-specific components of DETR with self-supervision from additional object proposals generated using Selective Search as pseudo labels. We further introduce an incremental few-shot fine-tuning strategy with knowledge distillation on the class-specific components of DETR to encourage the network in detecting novel classes without forgetting the base classes. Extensive experiments conducted on standard incremental object detection and incremental few-shot object detection settings show that our approach significantly outperforms state-of-the-art methods by a large margin.
Convolutional neural networks are prevailing in deep learning tasks. However, they suffer from massive cost issues when working on mobile devices. Network pruning is an effective method of model compression to handle such problems. This paper presents a novel structured network pruning method with auxiliary gating structures which assigns importance marks to blocks in backbone network as a criterion when pruning. Block-wise pruning is then realized by proposed voting strategy, which is different from prevailing methods who prune a model in small granularity like channel-wise. We further develop a three-stage training scheduling for the proposed architecture incorporating knowledge distillation for better performance. Our experiments demonstrate that our method can achieve state-of-the-arts compression performance for the classification tasks. In addition, our approach can integrate synergistically with other pruning methods by providing pretrained models, thus achieving a better performance than the unpruned model with over 93\% FLOPs reduced.
Concept-based explanations aims to fill the model interpretability gap for non-technical humans-in-the-loop. Previous work has focused on providing concepts for specific models (eg, neural networks) or data types (eg, images), and by either trying to extract concepts from an already trained network or training self-explainable models through multi-task learning. In this work, we propose ConceptDistil, a method to bring concept explanations to any black-box classifier using knowledge distillation. ConceptDistil is decomposed into two components:(1) a concept model that predicts which domain concepts are present in a given instance, and (2) a distillation model that tries to mimic the predictions of a black-box model using the concept model predictions. We validate ConceptDistil in a real world use-case, showing that it is able to optimize both tasks, bringing concept-explainability to any black-box model.