Knowledge distillation is an effective approach to leverage a well-trained network or an ensemble of them, named as the teacher, to guide the training of a student network. The outputs from the teacher network are used as soft labels for supervising the training of a new network. Recent studies \citep{muller2019does,yuan2020revisiting} revealed an intriguing property of the soft labels that making labels soft serves as a good regularization to the student network. From the perspective of statistical learning, regularization aims to reduce the variance, however how bias and variance change is not clear for training with soft labels. In this paper, we investigate the bias-variance tradeoff brought by distillation with soft labels. Specifically, we observe that during training the bias-variance tradeoff varies sample-wisely. Further, under the same distillation temperature setting, we observe that the distillation performance is negatively associated with the number of some specific samples, which are named as regularization samples since these samples lead to bias increasing and variance decreasing. Nevertheless, we empirically find that completely filtering out regularization samples also deteriorates distillation performance. Our discoveries inspired us to propose the novel weighted soft labels to help the network adaptively handle the sample-wise bias-variance tradeoff. Experiments on standard evaluation benchmarks validate the effectiveness of our method. Our code is available at \url{https://github.com/bellymonster/Weighted-Soft-Label-Distillation}.
Generative Adversarial Networks (GAN) is an adversarial model, and it has been demonstrated to be effective for various generative tasks. However, GAN and its variants also suffer from many training problems, such as mode collapse and gradient vanish. In this paper, we firstly propose a general crossover operator, which can be widely applied to GANs using evolutionary strategies. Then we design an evolutionary GAN framework C-GAN based on it. And we combine the crossover operator with evolutionary generative adversarial networks (EGAN) to implement the evolutionary generative adversarial networks with crossover (CE-GAN). Under the premise that a variety of loss functions are used as mutation operators to generate mutation individuals, we evaluate the generated samples and allow the mutation individuals to learn experiences from the output in a knowledge distillation manner, imitating the best output outcome, resulting in better offspring. Then, we greedily selected the best offspring as parents for subsequent training using discriminator as evaluator. Experiments on real datasets demonstrate the effectiveness of CE-GAN and show that our method is competitive in terms of generated images quality and time efficiency.
An accurate and efficient forecasting system is imperative to the prevention of emerging infectious diseases such as COVID-19 in public health. This system requires accurate transient modeling, lower computation cost, and fewer observation data. To tackle these three challenges, we propose a novel deep learning approach using black-box knowledge distillation for both accurate and efficient transmission dynamics prediction in a practical manner. First, we leverage mixture models to develop an accurate, comprehensive, yet impractical simulation system. Next, we use simulated observation sequences to query the simulation system to retrieve simulated projection sequences as knowledge. Then, with the obtained query data, sequence mixup is proposed to improve query efficiency, increase knowledge diversity, and boost distillation model accuracy. Finally, we train a student deep neural network with the retrieved and mixed observation-projection sequences for practical use. The case study on COVID-19 justifies that our approach accurately projects infections with much lower computation cost when observation data are limited.
Conventional transfer learning leverages weights of pre-trained networks, but mandates the need for similar neural architectures. Alternatively, knowledge distillation can transfer knowledge between heterogeneous networks but often requires access to the original training data or additional generative networks. Knowledge transfer between networks can be improved by being agnostic to the choice of network architecture and reducing the dependence on original training data. We propose a knowledge transfer approach from a teacher to a student network wherein we train the student on an independent transferal dataset, whose annotations are generated by the teacher. Experiments were conducted on five state-of-the-art networks for semantic segmentation and seven datasets across three imaging modalities. We studied knowledge transfer from a single teacher, combination of knowledge transfer and fine-tuning, and knowledge transfer from multiple teachers. The student model with a single teacher achieved similar performance as the teacher; and the student model with multiple teachers achieved better performance than the teachers. The salient features of our algorithm include: 1)no need for original training data or generative networks, 2) knowledge transfer between different architectures, 3) ease of implementation for downstream tasks by using the downstream task dataset as the transferal dataset, 4) knowledge transfer of an ensemble of models, trained independently, into one student model. Extensive experiments demonstrate that the proposed algorithm is effective for knowledge transfer and easily tunable.
Rehearsal is a critical component for class-incremental continual learning, yet it requires a substantial memory budget. Our work investigates whether we can significantly reduce this memory budget by leveraging unlabeled data from an agent's environment in a realistic and challenging continual learning paradigm. Specifically, we explore and formalize a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce yet non-i.i.d. unlabeled data from the agent's environment is plentiful. Importantly, data distributions in the SSCL setting are realistic and therefore reflect object class correlations between, and among, the labeled and unlabeled data distributions. We show that a strategy built on pseudo-labeling, consistency regularization, Out-of-Distribution (OoD) detection, and knowledge distillation reduces forgetting in this setting. Our approach, DistillMatch, increases performance over the state-of-the-art by no less than 8.7% average task accuracy and up to 54.5% average task accuracy in SSCL CIFAR-100 experiments. Moreover, we demonstrate that DistillMatch can save up to 0.23 stored images per processed unlabeled image compared to the next best method which only saves 0.08. Our results suggest that focusing on realistic correlated distributions is a significantly new perspective, which accentuates the importance of leveraging the world's structure as a continual learning strategy.
Nowadays, many of the images captured are `observed' by machines only and not by humans, e.g., in autonomous systems. High-level machine vision models, such as object recognition or semantic segmentation, assume images are transformed into some canonical image space by the camera \ans{Image Signal Processor (ISP)}. However, the camera ISP is optimized for producing visually pleasing images for human observers and not for machines. Therefore, one may spare the ISP compute time and apply vision models directly to RAW images. Yet, it has been shown that training such models directly on RAW images results in a performance drop. To mitigate this drop, we use a RAW and RGB image pairs dataset, which can be easily acquired with no human labeling. We then train a model that is applied directly to the RAW data by using knowledge distillation such that the model predictions for RAW images will be aligned with the predictions of an off-the-shelf pre-trained model for processed RGB images. Our experiments show that our performance on RAW images for object classification and semantic segmentation is significantly better than models trained on labeled RAW images. It also reasonably matches the predictions of a pre-trained model on processed RGB images, while saving the ISP compute overhead.
Action recognition, early prediction, and online action detection are complementary disciplines that are often studied independently. Most online action detection networks use a pre-trained feature extractor, which might not be optimal for its new task. We address the task-specific feature extraction with a teacher-student framework between the aforementioned disciplines, and a novel training strategy. Our network, Online Knowledge Distillation Action Detection network (OKDAD), embeds online early prediction and online temporal segment proposal subnetworks in parallel. Low interclass and high intraclass similarity are encouraged during teacher training. Knowledge distillation to the OKDAD network is ensured via layer reuse and cosine similarity between teacher-student feature vectors. Layer reuse and similarity learning significantly improve our baseline which uses a generic feature extractor. We evaluate our framework on infrared videos from two popular datasets, NTU RGB+D (action recognition, early prediction) and PKU MMD (action detection). Unlike previous attempts on those datasets, our student networks perform without any knowledge of the future. Even with this added difficulty, we achieve state-of-the-art results on both datasets. Moreover, our networks use infrared from RGB-D cameras, which we are the first to use for online action detection, to our knowledge.
Knowledge distillation (KD), as an efficient and effective model compression technique, has been receiving considerable attention in deep learning. The key to its success is to transfer knowledge from a large teacher network to a small student one. However, most of the existing knowledge distillation methods consider only one type of knowledge learned from either instance features or instance relations via a specific distillation strategy in teacher-student learning. There are few works that explore the idea of transferring different types of knowledge with different distillation strategies in a unified framework. Moreover, the frequently used offline distillation suffers from a limited learning capacity due to the fixed teacher-student architecture. In this paper we propose a collaborative teacher-student learning via multiple knowledge transfer (CTSL-MKT) that prompts both self-learning and collaborative learning. It allows multiple students learn knowledge from both individual instances and instance relations in a collaborative way. While learning from themselves with self-distillation, they can also guide each other via online distillation. The experiments and ablation studies on four image datasets demonstrate that the proposed CTSL-MKT significantly outperforms the state-of-the-art KD methods.
Despite pre-trained language models such as BERT have achieved appealing performance in a wide range of natural language processing tasks, they are computationally expensive to be deployed in real-time applications. A typical method is to adopt knowledge distillation to compress these large pre-trained models (teacher models) to small student models. However, for a target domain with scarce training data, the teacher can hardly pass useful knowledge to the student, which yields performance degradation for the student models. To tackle this problem, we propose a method to learn to augment for data-scarce domain BERT knowledge distillation, by learning a cross-domain manipulation scheme that automatically augments the target with the help of resource-rich source domains. Specifically, the proposed method generates samples acquired from a stationary distribution near the target data and adopts a reinforced selector to automatically refine the augmentation strategy according to the performance of the student. Extensive experiments demonstrate that the proposed method significantly outperforms state-of-the-art baselines on four different tasks, and for the data-scarce domains, the compressed student models even perform better than the original large teacher model, with much fewer parameters (only ${\sim}13.3\%$) when only a few labeled examples available.
Beyond the complexity of CNNs that require training on large annotated datasets, the domain shift between design and operational data has limited the adoption of CNNs in many real-world applications. For instance, in person re-identification, videos are captured over a distributed set of cameras with non-overlapping viewpoints. The shift between the source (e.g. lab setting) and target (e.g. cameras) domains may lead to a significant decline in recognition accuracy. Additionally, state-of-the-art CNNs may not be suitable for such real-time applications given their computational requirements. Although several techniques have recently been proposed to address domain shift problems through unsupervised domain adaptation (UDA), or to accelerate/compress CNNs through knowledge distillation (KD), we seek to simultaneously adapt and compress CNNs to generalize well across multiple target domains. In this paper, we propose a progressive KD approach for unsupervised single-target DA (STDA) and multi-target DA (MTDA) of CNNs. Our method for KD-STDA adapts a CNN to a single target domain by distilling from a larger teacher CNN, trained on both target and source domain data in order to maintain its consistency with a common representation. Our proposed approach is compared against state-of-the-art methods for compression and STDA of CNNs on the Office31 and ImageClef-DA image classification datasets. It is also compared against state-of-the-art methods for MTDA on Digits, Office31, and OfficeHome. In both settings -- KD-STDA and KD-MTDA -- results indicate that our approach can achieve the highest level of accuracy across target domains, while requiring a comparable or lower CNN complexity.
In the past years, Knowledge-Based Question Answering (KBQA), which aims to answer natural language questions using facts in a knowledge base, has been well developed. Existing approaches often assume a static knowledge base. However, the knowledge is evolving over time in the real world. If we directly apply a fine-tuning strategy on an evolving knowledge base, it will suffer from a serious catastrophic forgetting problem. In this paper, we propose a new incremental KBQA learning framework that can progressively expand learning capacity as humans do. Specifically, it comprises a margin-distilled loss and a collaborative exemplar selection method, to overcome the catastrophic forgetting problem by taking advantage of knowledge distillation. We reorganize the SimpleQuestion dataset to evaluate the proposed incremental learning solution to KBQA. The comprehensive experiments demonstrate its effectiveness and efficiency when working with the evolving knowledge base.
Pretrained deep models hold their learnt knowledge in the form of model parameters. These parameters act as "memory" for the trained models and help them generalize well on unseen data. However, in absence of training data, the utility of a trained model is merely limited to either inference or better initialization towards a target task. In this paper, we go further and extract synthetic data by leveraging the learnt model parameters. We dub them "Data Impressions", which act as proxy to the training data and can be used to realize a variety of tasks. These are useful in scenarios where only the pretrained models are available and the training data is not shared (e.g., due to privacy or sensitivity concerns). We show the applicability of data impressions in solving several computer vision tasks such as unsupervised domain adaptation, continual learning as well as knowledge distillation. We also study the adversarial robustness of lightweight models trained via knowledge distillation using these data impressions. Further, we demonstrate the efficacy of data impressions in generating data-free Universal Adversarial Perturbations (UAPs) with better fooling rates. Extensive experiments performed on benchmark datasets demonstrate competitive performance achieved using data impressions in absence of original training data.
Recently, transformer-based language models such as BERT have shown tremendous performance improvement for a range of natural language processing tasks. However, these language models usually are computation expensive and memory intensive during inference. As a result, it is difficult to deploy them on resource-restricted devices. To improve the inference performance, as well as reduce the model size while maintaining the model accuracy, we propose a novel quantization method named KDLSQ-BERT that combines knowledge distillation (KD) with learned step size quantization (LSQ) for language model quantization. The main idea of our method is that the KD technique is leveraged to transfer the knowledge from a "teacher" model to a "student" model when exploiting LSQ to quantize that "student" model during the quantization training process. Extensive experiment results on GLUE benchmark and SQuAD demonstrate that our proposed KDLSQ-BERT not only performs effectively when doing different bit (e.g. 2-bit $\sim$ 8-bit) quantization, but also outperforms the existing BERT quantization methods, and even achieves comparable performance as the full-precision base-line model while obtaining 14.9x compression ratio. Our code will be public available.
On-device Deep Neural Networks (DNNs) have recently gained more attention due to the increasing computing power of the mobile devices and the number of applications in Computer Vision (CV), Natural Language Processing (NLP), and Internet of Things (IoTs). Unfortunately, the existing efficient convolutional neural network (CNN) architectures designed for CV tasks are not directly applicable to NLP tasks and the tiny Recurrent Neural Network (RNN) architectures have been designed primarily for IoT applications. In NLP applications, although model compression has seen initial success in on-device text classification, there are at least three major challenges yet to be addressed: adversarial robustness, explainability, and personalization. Here we attempt to tackle these challenges by designing a new training scheme for model compression and adversarial robustness, including the optimization of an explainable feature mapping objective, a knowledge distillation objective, and an adversarially robustness objective. The resulting compressed model is personalized using on-device private training data via fine-tuning. We perform extensive experiments to compare our approach with both compact RNN (e.g., FastGRNN) and compressed RNN (e.g., PRADO) architectures in both natural and adversarial NLP test settings.
We propose a novel framework for producing a class of parameter and compute efficient models called AttentionLitesuitable for resource-constrained applications. Prior work has primarily focused on optimizing models either via knowledge distillation or pruning. In addition to fusing these two mechanisms, our joint optimization framework also leverages recent advances in self-attention as a substitute for convolutions. We can simultaneously distill knowledge from a compute-heavy teacher while also pruning the student model in a single pass of training thereby reducing training and fine-tuning times considerably. We evaluate the merits of our proposed approach on the CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. Not only do our AttentionLite models significantly outperform their unoptimized counterparts in accuracy, we find that in some cases, that they perform almost as well as their compute-heavy teachers while consuming only a fraction of the parameters and FLOPs. Concretely, AttentionLite models can achieve upto30x parameter efficiency and 2x computation efficiency with no significant accuracy drop compared to their teacher.
This paper is concerned with self-supervised learning for small models. The problem is motivated by our empirical studies that while the widely used contrastive self-supervised learning method has shown great progress on large model training, it does not work well for small models. To address this problem, we propose a new learning paradigm, named SElf-SupErvised Distillation (SEED), where we leverage a larger network (as Teacher) to transfer its representational knowledge into a smaller architecture (as Student) in a self-supervised fashion. Instead of directly learning from unlabeled data, we train a student encoder to mimic the similarity score distribution inferred by a teacher over a set of instances. We show that SEED dramatically boosts the performance of small networks on downstream tasks. Compared with self-supervised baselines, SEED improves the top-1 accuracy from 42.2% to 67.6% on EfficientNet-B0 and from 36.3% to 68.2% on MobileNet-v3-Large on the ImageNet-1k dataset.
Machine learning models of materials$^{1-5}$ accelerate discovery compared to ab initio methods: deep learning models now reproduce density functional theory (DFT)-calculated results at one hundred thousandths of the cost of DFT$^{6}$. To provide guidance in experimental materials synthesis, these need to be coupled with an accurate yet effective search algorithm and training data consistent with experimental observations. Here we report an evolutionary algorithm powered search which uses machine-learned surrogate models trained on high-throughput hybrid functional DFT data benchmarked against experimental bandgaps: Deep Adaptive Regressive Weighted Intelligent Network (DARWIN). The strategy enables efficient search over the materials space of ~10$^8$ ternaries and 10$^{11}$ quaternaries$^{7}$ for candidates with target properties. It provides interpretable design rules, such as our finding that the difference in the electronegativity between the halide and B-site cation being a strong predictor of ternary structural stability. As an example, when we seek UV emission, DARWIN predicts K$_2$CuX$_3$ (X = Cl, Br) as a promising materials family, based on its electronegativity difference. We synthesized and found these materials to be stable, direct bandgap UV emitters. The approach also allows knowledge distillation for use by humans.
Developing deep learning models to analyze histology images has been computationally challenging, as the massive size of the images causes excessive strain on all parts of the computing pipeline. This paper proposes a novel deep learning-based methodology for improving the computational efficiency of histology image classification. The proposed approach is robust when used with images that have reduced input resolution and can be trained effectively with limited labeled data. Pre-trained on the original high-resolution (HR) images, our method uses knowledge distillation (KD) to transfer learned knowledge from a teacher model to a student model trained on the same images at a much lower resolution. To address the lack of large-scale labeled histology image datasets, we perform KD in a self-supervised manner. We evaluate our approach on two histology image datasets associated with celiac disease (CD) and lung adenocarcinoma (LUAD). Our results show that a combination of KD and self-supervision allows the student model to approach, and in some cases, surpass the classification accuracy of the teacher, while being much more efficient. Additionally, we observe an increase in student classification performance as the size of the unlabeled dataset increases, indicating that there is potential to scale further. For the CD data, our model outperforms the HR teacher model, while needing 4 times fewer computations. For the LUAD data, our student model results at 1.25x magnification are within 3% of the teacher model at 10x magnification, with a 64 times computational cost reduction. Moreover, our CD outcomes benefit from performance scaling with the use of more unlabeled data. For 0.625x magnification, using unlabeled data improves accuracy by 4% over the baseline. Thus, our method can improve the feasibility of deep learning solutions for digital pathology with standard computational hardware.
Lottery Ticket Hypothesis (LTH) suggests that a dense neural network contains a sparse sub-network that can match the performance of the original dense network when trained in isolation from scratch. Most works retrain the sparse sub-network with the same training protocols as its dense network, such as initialization, architecture blocks, and training recipes. However, till now it is unclear that whether these training protocols are optimal for sparse networks. In this paper, we demonstrate that it is unnecessary for spare retraining to strictly inherit those properties from the dense network. Instead, by plugging in purposeful "tweaks" of the sparse subnetwork architecture or its training recipe, its retraining can be significantly improved than the default, especially at high sparsity levels. Combining all our proposed "tweaks" can yield the new state-of-the-art performance of LTH, and these modifications can be easily adapted to other sparse training algorithms in general. Specifically, we have achieved a significant and consistent performance gain of1.05% - 4.93% for ResNet18 on CIFAR-100 over vanilla-LTH. Moreover, our methods are shown to generalize across datasets (CIFAR10, CIFAR100, TinyImageNet) and architectures (Vgg16, ResNet-18/ResNet-34, MobileNet). All codes will be publicly available.
Iterative generative models, such as noise conditional score networks and denoising diffusion probabilistic models, produce high quality samples by gradually denoising an initial noise vector. However, their denoising process has many steps, making them 2-3 orders of magnitude slower than other generative models such as GANs and VAEs. In this paper, we establish a novel connection between knowledge distillation and image generation with a technique that distills a multi-step denoising process into a single step, resulting in a sampling speed similar to other single-step generative models. Our Denoising Student generates high quality samples comparable to GANs on the CIFAR-10 and CelebA datasets, without adversarial training. We demonstrate that our method scales to higher resolutions through experiments on 256 x 256 LSUN. Code and checkpoints are available at https://github.com/tcl9876/Denoising_Student