Structured prediction models aim at solving a type of problem where the output is a complex structure, rather than a single variable. Performing knowledge distillation for such models is not trivial due to their exponentially large output space. In this work, we propose an approach that is much simpler in its formulation and far more efficient for training than existing approaches. Specifically, we transfer the knowledge from a teacher model to its student model by locally matching their predictions on all sub-structures, instead of the whole output space. In this manner, we avoid adopting some time-consuming techniques like dynamic programming (DP) for decoding output structures, which permits parallel computation and makes the training process even faster in practice. Besides, it encourages the student model to better mimic the internal behavior of the teacher model. Experiments on two structured prediction tasks demonstrate that our approach outperforms previous methods and halves the time cost for one training epoch.
Deep learning-based image signal processor (ISP) models for mobile cameras can generate high-quality images that rival those of professional DSLR cameras. However, their computational demands often make them unsuitable for mobile settings. Additionally, modern mobile cameras employ non-Bayer color filter arrays (CFA) such as Quad Bayer, Nona Bayer, and QxQ Bayer to enhance image quality, yet most existing deep learning-based ISP (or demosaicing) models focus primarily on standard Bayer CFAs. In this study, we present PyNET-QxQ, a lightweight demosaicing model specifically designed for QxQ Bayer CFA patterns, which is derived from the original PyNET. We also propose a knowledge distillation method called progressive distillation to train the reduced network more effectively. Consequently, PyNET-QxQ contains less than 2.5% of the parameters of the original PyNET while preserving its performance. Experiments using QxQ images captured by a proto type QxQ camera sensor show that PyNET-QxQ outperforms existing conventional algorithms in terms of texture and edge reconstruction, despite its significantly reduced parameter count.
Unsupervised domain adaptation has been proposed recently to tackle the so-called domain shift between training data and test data with different distributions. However, most of them only focus on single-target domain adaptation and cannot be applied to the scenario with multiple target domains. In this paper, we propose RVms, a novel unsupervised multi-target domain adaptation approach to segment retinal vessels (RVs) from multimodal and multicenter retinal images. RVms mainly consists of a style augmentation and transfer (SAT) module and a dual-teacher knowledge distillation (DTKD) module. SAT augments and clusters images into source-similar domains and source-dissimilar domains via Bezier and Fourier transformations. DTKD utilizes the augmented and transformed data to train two teachers, one for source-similar domains and the other for source-dissimilar domains. Afterwards, knowledge distillation is performed to iteratively distill different domain knowledge from teachers to a generic student. The local relative intensity transformation is employed to characterize RVs in a domain invariant manner and promote the generalizability of teachers and student models. Moreover, we construct a new multimodal and multicenter vascular segmentation dataset from existing publicly-available datasets, which can be used to benchmark various domain adaptation and domain generalization methods. Through extensive experiments, RVms is found to be very close to the target-trained Oracle in terms of segmenting the RVs, largely outperforming other state-of-the-art methods.
Neural networks tend to gradually forget the previously learned knowledge when learning multiple tasks sequentially from dynamic data distributions. This problem is called \textit{catastrophic forgetting}, which is a fundamental challenge in the continual learning of neural networks. In this work, we observe that catastrophic forgetting not only occurs in continual learning but also affects the traditional static training. Neural networks, especially neural machine translation models, suffer from catastrophic forgetting even if they learn from a static training set. To be specific, the final model pays imbalanced attention to training samples, where recently exposed samples attract more attention than earlier samples. The underlying cause is that training samples do not get balanced training in each model update, so we name this problem \textit{imbalanced training}. To alleviate this problem, we propose Complementary Online Knowledge Distillation (COKD), which uses dynamically updated teacher models trained on specific data orders to iteratively provide complementary knowledge to the student model. Experimental results on multiple machine translation tasks show that our method successfully alleviates the problem of imbalanced training and achieves substantial improvements over strong baseline systems.
In this paper, we propose to employ a dual-mode framework on the x-vector self-attention (XSA-LID) model with knowledge distillation (KD) to enhance its language identification (LID) performance for both long and short utterances. The dual-mode XSA-LID model is trained by jointly optimizing both the full and short modes with their respective inputs being the full-length speech and its short clip extracted by a specific Boolean mask, and KD is applied to further boost the performance on short utterances. In addition, we investigate the impact of clip-wise linguistic variability and lexical integrity for LID by analyzing the variation of LID performance in terms of the lengths and positions of the mimicked speech clips. We evaluated our approach on the MLS14 data from the NIST 2017 LRE. With the 3~s random-location Boolean mask, our proposed method achieved 19.23%, 21.52% and 8.37% relative improvement in average cost compared with the XSA-LID model on 3s, 10s, and 30s speech, respectively.
Continual relation extraction (CRE) aims to continuously train a model on data with new relations while avoiding forgetting old ones. Some previous work has proved that storing a few typical samples of old relations and replaying them when learning new relations can effectively avoid forgetting. However, these memory-based methods tend to overfit the memory samples and perform poorly on imbalanced datasets. To solve these challenges, a consistent representation learning method is proposed, which maintains the stability of the relation embedding by adopting contrastive learning and knowledge distillation when replaying memory. Specifically, supervised contrastive learning based on a memory bank is first used to train each new task so that the model can effectively learn the relation representation. Then, contrastive replay is conducted of the samples in memory and makes the model retain the knowledge of historical relations through memory knowledge distillation to prevent the catastrophic forgetting of the old task. The proposed method can better learn consistent representations to alleviate forgetting effectively. Extensive experiments on FewRel and TACRED datasets show that our method significantly outperforms state-of-the-art baselines and yield strong robustness on the imbalanced dataset.
Conventional NAS-based pruning algorithms aim to find the sub-network with the best validation performance. However, validation performance does not successfully represent test performance, i.e., potential performance. Also, although fine-tuning the pruned network to restore the performance drop is an inevitable process, few studies have handled this issue. This paper provides a novel Ensemble Knowledge Guidance (EKG) to solve both problems at once. First, we experimentally prove that the fluctuation of loss landscape can be an effective metric to evaluate the potential performance. In order to search a sub-network with the smoothest loss landscape at a low cost, we employ EKG as a search reward. EKG utilized for the following search iteration is composed of the ensemble knowledge of interim sub-networks, i.e., the by-products of the sub-network evaluation. Next, we reuse EKG to provide a gentle and informative guidance to the pruned network while fine-tuning the pruned network. Since EKG is implemented as a memory bank in both phases, it requires a negligible cost. For example, when pruning and training ResNet-50, just 315 GPU hours are required to remove around 45.04% of FLOPS without any performance degradation, which can operate even on a low-spec workstation. the implemented code is available at https://github.com/sseung0703/EKG.
Better-supervised models might have better performance. In this paper, we first clarify what makes for good supervision for a classification problem, and then explain two existing label refining methods, label smoothing and knowledge distillation, in terms of our proposed criterion. To further answer why and how better supervision emerges, we observe the learning path, i.e., the trajectory of the model's predictions during training, for each training sample. We find that the model can spontaneously refine "bad" labels through a "zig-zag" learning path, which occurs on both toy and real datasets. Observing the learning path not only provides a new perspective for understanding knowledge distillation, overfitting, and learning dynamics, but also reveals that the supervisory signal of a teacher network can be very unstable near the best points in training on real tasks. Inspired by this, we propose a new knowledge distillation scheme, Filter-KD, which improves downstream classification performance in various settings.
3D dense captioning aims to describe individual objects by natural language in 3D scenes, where 3D scenes are usually represented as RGB-D scans or point clouds. However, only exploiting single modal information, e.g., point cloud, previous approaches fail to produce faithful descriptions. Though aggregating 2D features into point clouds may be beneficial, it introduces an extra computational burden, especially in inference phases. In this study, we investigate a cross-modal knowledge transfer using Transformer for 3D dense captioning, X-Trans2Cap, to effectively boost the performance of single-modal 3D caption through knowledge distillation using a teacher-student framework. In practice, during the training phase, the teacher network exploits auxiliary 2D modality and guides the student network that only takes point clouds as input through the feature consistency constraints. Owing to the well-designed cross-modal feature fusion module and the feature alignment in the training phase, X-Trans2Cap acquires rich appearance information embedded in 2D images with ease. Thus, a more faithful caption can be generated only using point clouds during the inference. Qualitative and quantitative results confirm that X-Trans2Cap outperforms previous state-of-the-art by a large margin, i.e., about +21 and about +16 absolute CIDEr score on ScanRefer and Nr3D datasets, respectively.
Tissue phenotyping is a fundamental task in learning objective characterizations of histopathologic biomarkers within the tumor-immune microenvironment in cancer pathology. However, whole-slide imaging (WSI) is a complex computer vision in which: 1) WSIs have enormous image resolutions with precludes large-scale pixel-level efforts in data curation, and 2) diversity of morphological phenotypes results in inter- and intra-observer variability in tissue labeling. To address these limitations, current efforts have proposed using pretrained image encoders (transfer learning from ImageNet, self-supervised pretraining) in extracting morphological features from pathology, but have not been extensively validated. In this work, we conduct a search for good representations in pathology by training a variety of self-supervised models with validation on a variety of weakly-supervised and patch-level tasks. Our key finding is in discovering that Vision Transformers using DINO-based knowledge distillation are able to learn data-efficient and interpretable features in histology images wherein the different attention heads learn distinct morphological phenotypes. We make evaluation code and pretrained weights publicly-available at: https://github.com/Richarizardd/Self-Supervised-ViT-Path.
Semantic segmentation benchmarks in the realm of autonomous driving are dominated by large pre-trained transformers, yet their widespread adoption is impeded by substantial computational costs and prolonged training durations. To lift this constraint, we look at efficient semantic segmentation from a perspective of comprehensive knowledge distillation and aim to bridge the gap between multi-source knowledge extractions and transformer-specific patch embeddings. We put forward the Transformer-based Knowledge Distillation (TransKD) framework which learns compact student transformers by distilling both feature maps and patch embeddings of large teacher transformers, bypassing the long pre-training process and reducing the FLOPs by >85.0%. Specifically, we propose two fundamental modules to realize feature map distillation and patch embedding distillation, respectively: (1) Cross Selective Fusion (CSF) enables knowledge transfer between cross-stage features via channel attention and feature map distillation within hierarchical transformers; (2) Patch Embedding Alignment (PEA) performs dimensional transformation within the patchifying process to facilitate the patch embedding distillation. Furthermore, we introduce two optimization modules to enhance the patch embedding distillation from different perspectives: (1) Global-Local Context Mixer (GL-Mixer) extracts both global and local information of a representative embedding; (2) Embedding Assistant (EA) acts as an embedding method to seamlessly bridge teacher and student models with the teacher's number of channels. Experiments on Cityscapes, ACDC, NYUv2, and Pascal VOC2012 datasets show that TransKD outperforms state-of-the-art distillation frameworks and rivals the time-consuming pre-training method. The source code is publicly available at https://github.com/RuipingL/TransKD.
Generally, humans are more skilled at perceiving differences between high-quality (HQ) and low-quality (LQ) images than directly judging the quality of a single LQ image. This situation also applies to image quality assessment (IQA). Although recent no-reference (NR-IQA) methods have made great progress to predict image quality free from the reference image, they still have the potential to achieve better performance since HQ image information is not fully exploited. In contrast, full-reference (FR-IQA) methods tend to provide more reliable quality evaluation, but its practicability is affected by the requirement for pixel-level aligned reference images. To address this, we firstly propose the content-variant reference method via knowledge distillation (CVRKD-IQA). Specifically, we use non-aligned reference (NAR) images to introduce various prior distributions of high-quality images. The comparisons of distribution differences between HQ and LQ images can help our model better assess the image quality. Further, the knowledge distillation transfers more HQ-LQ distribution difference information from the FR-teacher to the NAR-student and stabilizing CVRKD-IQA performance. Moreover, to fully mine the local-global combined information, while achieving faster inference speed, our model directly processes multiple image patches from the input with the MLP-mixer. Cross-dataset experiments verify that our model can outperform all NAR/NR-IQA SOTAs, even reach comparable performance with FR-IQA methods on some occasions. Since the content-variant and non-aligned reference HQ images are easy to obtain, our model can support more IQA applications with its relative robustness to content variations. Our code and more detailed elaborations of supplements are available: https://github.com/guanghaoyin/CVRKD-IQA.
Visual Commonsense Reasoning (VCR), deemed as one challenging extension of the Visual Question Answering (VQA), endeavors to pursue a more high-level visual comprehension. It is composed of two indispensable processes: question answering over a given image and rationale inference for answer explanation. Over the years, a variety of methods tackling VCR have advanced the performance on the benchmark dataset. Despite significant as these methods are, they often treat the two processes in a separate manner and hence decompose the VCR into two irrelevant VQA instances. As a result, the pivotal connection between question answering and rationale inference is interrupted, rendering existing efforts less faithful on visual reasoning. To empirically study this issue, we perform some in-depth explorations in terms of both language shortcuts and generalization capability to verify the pitfalls of this treatment. Based on our findings, in this paper, we present a plug-and-play knowledge distillation enhanced framework to couple the question answering and rationale inference processes. The key contribution is the introduction of a novel branch, which serves as the bridge to conduct processes connecting. Given that our framework is model-agnostic, we apply it to the existing popular baselines and validate its effectiveness on the benchmark dataset. As detailed in the experimental results, when equipped with our framework, these baselines achieve consistent and significant performance improvements, demonstrating the viability of processes coupling, as well as the superiority of the proposed framework.
The number of linear regions has been studied as a proxy of complexity for ReLU networks. However, the empirical success of network compression techniques like pruning and knowledge distillation, suggest that in the overparameterized setting, linear regions density might fail to capture the effective nonlinearity. In this work, we propose an efficient algorithm for discovering linear regions and use it to investigate the effectiveness of density in capturing the nonlinearity of trained VGGs and ResNets on CIFAR-10 and CIFAR-100. We contrast the results with a more principled nonlinearity measure based on function variation, highlighting the shortcomings of linear regions density. Furthermore, interestingly, our measure of nonlinearity clearly correlates with model-wise deep double descent, connecting reduced test error with reduced nonlinearity, and increased local similarity of linear regions.
Describing images in natural language is a fundamental step towards the automatic modeling of connections between the visual and textual modalities. In this paper we present CaMEL, a novel Transformer-based architecture for image captioning. Our proposed approach leverages the interaction of two interconnected language models that learn from each other during the training phase. The interplay between the two language models follows a mean teacher learning paradigm with knowledge distillation. Experimentally, we assess the effectiveness of the proposed solution on the COCO dataset and in conjunction with different visual feature extractors. When comparing with existing proposals, we demonstrate that our model provides state-of-the-art caption quality with a significantly reduced number of parameters. According to the CIDEr metric, we obtain a new state of the art on COCO when training without using external data. The source code and trained models are publicly available at: https://github.com/aimagelab/camel.
Recent studies in Learning to Rank have shown the possibility to effectively distill a neural network from an ensemble of regression trees. This result leads neural networks to become a natural competitor of tree-based ensembles on the ranking task. Nevertheless, ensembles of regression trees outperform neural models both in terms of efficiency and effectiveness, particularly when scoring on CPU. In this paper, we propose an approach for speeding up neural scoring time by applying a combination of Distillation, Pruning and Fast Matrix multiplication. We employ knowledge distillation to learn shallow neural networks from an ensemble of regression trees. Then, we exploit an efficiency-oriented pruning technique that performs a sparsification of the most computationally-intensive layers of the neural network that is then scored with optimized sparse matrix multiplication. Moreover, by studying both dense and sparse high performance matrix multiplication, we develop a scoring time prediction model which helps in devising neural network architectures that match the desired efficiency requirements. Comprehensive experiments on two public learning-to-rank datasets show that neural networks produced with our novel approach are competitive at any point of the effectiveness-efficiency trade-off when compared with tree-based ensembles, providing up to 4x scoring time speed-up without affecting the ranking quality.
This paper describes the principle of "General Cyclical Training" in machine learning, where training starts and ends with "easy training" and the "hard training" happens during the middle epochs. We propose several manifestations for training neural networks, including algorithmic examples (via hyper-parameters and loss functions), data-based examples, and model-based examples. Specifically, we introduce several novel techniques: cyclical weight decay, cyclical batch size, cyclical focal loss, cyclical softmax temperature, cyclical data augmentation, cyclical gradient clipping, and cyclical semi-supervised learning. In addition, we demonstrate that cyclical weight decay, cyclical softmax temperature, and cyclical gradient clipping (as three examples of this principle) are beneficial in the test accuracy performance of a trained model. Furthermore, we discuss model-based examples (such as pretraining and knowledge distillation) from the perspective of general cyclical training and recommend some changes to the typical training methodology. In summary, this paper defines the general cyclical training concept and discusses several specific ways in which this concept can be applied to training neural networks. In the spirit of reproducibility, the code used in our experiments is available at \url{https://github.com/lnsmith54/CFL}.
This paper presents FAMIE, a comprehensive and efficient active learning (AL) toolkit for multilingual information extraction. FAMIE is designed to address a fundamental problem in existing AL frameworks where annotators need to wait for a long time between annotation batches due to the time-consuming nature of model training and data selection at each AL iteration. This hinders the engagement, productivity, and efficiency of annotators. Based on the idea of using a small proxy network for fast data selection, we introduce a novel knowledge distillation mechanism to synchronize the proxy network with the main large model (i.e., BERT-based) to ensure the appropriateness of the selected annotation examples for the main model. Our AL framework can support multiple languages. The experiments demonstrate the advantages of FAMIE in terms of competitive performance and time efficiency for sequence labeling with AL. We publicly release our code (\url{https://github.com/nlp-uoregon/famie}) and demo website (\url{http://nlp.uoregon.edu:9000/}). A demo video for FAMIE is provided at: \url{https://youtu.be/I2i8n_jAyrY}.
We introduce EdgeFormer -- a parameter-efficient Transformer for on-device seq2seq generation under the strict computation and memory constraints. Compared with the previous parameter-efficient Transformers, EdgeFormer applies two novel principles for cost-effective parameterization, allowing it to perform better given the same parameter budget; moreover, EdgeFormer is further enhanced by layer adaptation innovation that is proposed for improving the network with shared layers. Extensive experiments show EdgeFormer can effectively outperform previous parameter-efficient Transformer baselines and achieve competitive results under both the computation and memory constraints. Given the promising results, we release EdgeLM -- the pretrained version of EdgeFormer, which is the first publicly available pretrained on-device seq2seq model that can be easily fine-tuned for seq2seq tasks with strong results, facilitating on-device seq2seq generation in practice.
There is a growing interest in dataset generation recently due to the superior generative capacity of large pre-trained language models (PLMs). In this paper, we study a flexible and efficient zero-short learning method, \textsc{ZeroGen}. Given a zero-shot task, we first generate a dataset from scratch using PLMs in an unsupervised manner. Then, we train a tiny task model (e.g., LSTM) under the supervision of the synthesized dataset. This approach allows highly efficient inference as the final task model only has orders of magnitude fewer parameters comparing to PLMs (e.g., GPT2-XL). Apart from being annotation-free and efficient, we argue that \textsc{ZeroGen} can also provide useful insights from the perspective of data-free model-agnostic knowledge distillation, and unreferenced text generation evaluation. Experiments and analysis on different NLP tasks, namely, text classification, question answering, and natural language inference, show the effectiveness of \textsc{ZeroGen}.