In recent years, knowledge distillation (KD) has been widely used to derive efficient models. Through imitating a large teacher model, a lightweight student model can achieve comparable performance with more efficiency. However, most existing knowledge distillation methods are focused on classification tasks. Only a limited number of studies have applied knowledge distillation to object detection, especially in time-sensitive autonomous driving scenarios. In this paper, we propose Adaptive Instance Distillation (AID) to selectively impart teacher's knowledge to the student to improve the performance of knowledge distillation. Unlike previous KD methods that treat all instances equally, our AID can attentively adjust the distillation weights of instances based on the teacher model's prediction loss. We verified the effectiveness of our AID method through experiments on the KITTI and the COCO traffic datasets. The results show that our method improves the performance of state-of-the-art attention-guided and non-local distillation methods and achieves better distillation results on both single-stage and two-stage detectors. Compared to the baseline, our AID led to an average of 2.7% and 2.1% mAP increases for single-stage and two-stage detectors, respectively. Furthermore, our AID is also shown to be useful for self-distillation to improve the teacher model's performance.
Human education system trains one student by multiple experts. Mixture-of-experts (MoE) is a powerful sparse architecture including multiple experts. However, sparse MoE model is easy to overfit, hard to deploy, and not hardware-friendly for practitioners. In this work, inspired by the human education model, we propose a novel task, knowledge integration, to obtain a dense student model (OneS) as knowledgeable as one sparse MoE. We investigate this task by proposing a general training framework including knowledge gathering and knowledge distillation. Specifically, to gather key knowledge from different pre-trained experts, we first investigate four different possible knowledge gathering methods, \ie summation, averaging, Top-K Knowledge Gathering (Top-KG), and Singular Value Decomposition Knowledge Gathering (SVD-KG) proposed in this paper. We then refine the dense student model by knowledge distillation to offset the noise from gathering. On ImageNet, our OneS preserves $61.7\%$ benefits from MoE and achieves $78.4\%$ top-1 accuracy ImageNet with only $15$M parameters. On four natural language processing datasets, OneS obtains $88.2\%$ MoE benefits and outperforms the best baseline by $51.7\%$ using the same architecture and training data. In addition, compared with the MoE counterpart, OneS can achieve $3.7 \times$ inference speedup due to less computation and hardware-friendly architecture.
Knowledge distillation (KD) achieves promising results on the challenging problem of unsupervised anomaly detection (AD).The representation discrepancy of anomalies in the teacher-student (T-S) model provides essential evidence for AD. However, using similar or identical architectures to build the teacher and student models in previous studies hinders the diversity of anomalous representations. To tackle this problem, we propose a novel T-S model consisting of a teacher encoder and a student decoder and introduce a simple yet effective "reverse distillation" paradigm accordingly. Instead of receiving raw images directly, the student network takes teacher model's one-class embedding as input and targets to restore the teacher's multiscale representations. Inherently, knowledge distillation in this study starts from abstract, high-level presentations to low-level features. In addition, we introduce a trainable one-class bottleneck embedding (OCBE) module in our T-S model. The obtained compact embedding effectively preserves essential information on normal patterns, but abandons anomaly perturbations. Extensive experimentation on AD and one-class novelty detection benchmarks shows that our method surpasses SOTA performance, demonstrating our proposed approach's effectiveness and generalizability.
Multitask learning (MTL) has recently gained a lot of popularity as a learning paradigm that can lead to improved per-task performance while also using fewer per-task model parameters compared to single task learning. One of the biggest challenges regarding MTL networks involves how to share features across tasks. To address this challenge, we propose the Attentive Task Interaction Network (ATI-Net). ATI-Net employs knowledge distillation of the latent features for each task, then combines the feature maps to provide improved contextualized information to the decoder. This novel approach to introducing knowledge distillation into an attention based multitask network outperforms state of the art MTL baselines such as the standalone MTAN and PAD-Net, with roughly the same number of model parameters.
Federated Learning (FL) is designed to protect the data privacy of each client during the training process by transmitting only models instead of the original data. However, the trained model may memorize certain information about the training data. With the recent legislation on right to be forgotten, it is crucially essential for the FL model to possess the ability to forget what it has learned from each client. We propose a novel federated unlearning method to eliminate a client's contribution by subtracting the accumulated historical updates from the model and leveraging the knowledge distillation method to restore the model's performance without using any data from the clients. This method does not have any restrictions on the type of neural networks and does not rely on clients' participation, so it is practical and efficient in the FL system. We further introduce backdoor attacks in the training process to help evaluate the unlearning effect. Experiments on three canonical datasets demonstrate the effectiveness and efficiency of our method.
The gap in representations between image and video makes Image-to-Video Re-identification (I2V Re-ID) challenging, and recent works formulate this problem as a knowledge distillation (KD) process. In this paper, we propose a mutual discriminative knowledge distillation framework to transfer a video-based richer representation to an image based representation more effectively. Specifically, we propose the triplet contrast loss (TCL), a novel loss designed for KD. During the KD process, the TCL loss transfers the local structure, exploits the higher order information, and mitigates the misalignment of the heterogeneous output of teacher and student networks. Compared with other losses for KD, the proposed TCL loss selectively transfers the local discriminative features from teacher to student, making it effective in the ReID. Besides the TCL loss, we adopt mutual learning to regularize both the teacher and student networks training. Extensive experiments demonstrate the effectiveness of our method on the MARS, DukeMTMC-VideoReID and VeRi-776 benchmarks.
Model compression techniques are receiving increasing attention; however, the effect of compression on model fairness is still under explored. This is the first paper to examine the effect of distillation and pruning on the toxicity and bias of generative language models. We test Knowledge Distillation and Pruning methods on the GPT2 model and found a consistent pattern of toxicity and bias reduction after model distillation; this result can be potentially interpreted by existing line of research which describes model compression as a regularization technique; our work not only serves as a reference for safe deployment of compressed models, but also extends the discussion of "compression as regularization" into the setting of neural LMs, and hints at the possibility of using compression to develop fairer models.
Recently, large pre-trained models have significantly improved the performance of various Natural LanguageProcessing (NLP) tasks but they are expensive to serve due to long serving latency and large memory usage. To compress these models, knowledge distillation has attracted an increasing amount of interest as one of the most effective methods for model compression. However, existing distillation methods have not yet addressed the unique challenges of model serving in datacenters, such as handling fast evolving models, considering serving performance, and optimizing for multiple objectives. To solve these problems, we propose AutoDistill, an end-to-end model distillation framework integrating model architecture exploration and multi-objective optimization for building hardware-efficient NLP pre-trained models. We use Bayesian Optimization to conduct multi-objective Neural Architecture Search for selecting student model architectures. The proposed search comprehensively considers both prediction accuracy and serving latency on target hardware. The experiments on TPUv4i show the finding of seven model architectures with better pre-trained accuracy (up to 3.2% higher) and lower inference latency (up to 1.44x faster) than MobileBERT. By running downstream NLP tasks in the GLUE benchmark, the model distilled for pre-training by AutoDistill with 28.5M parameters achieves an 81.69 average score, which is higher than BERT_BASE, DistillBERT, TinyBERT, NAS-BERT, and MobileBERT. The most compact model found by AutoDistill contains only 20.6M parameters but still outperform BERT_BASE(109M), DistillBERT(67M), TinyBERT(67M), and MobileBERT(25.3M) regarding the average GLUE score. By evaluating on SQuAD, a model found by AutoDistill achieves an 88.4% F1 score with 22.8M parameters, which reduces parameters by more than 62% while maintaining higher accuracy than DistillBERT, TinyBERT, and NAS-BERT.
In online advertising, conventional post-click conversion rate (CVR) estimation models are trained using clicked samples. However, during online serving the models need to estimate for all impression ads, leading to the sample selection bias (SSB) issue. Intuitively, providing reliable supervision signals for unclicked ads is a feasible way to alleviate the SSB issue. This paper proposes an uncertainty-regularized knowledge distillation (UKD) framework to debias CVR estimation via distilling knowledge from unclicked ads. A teacher model learns click-adaptive representations and produces pseudo-conversion labels on unclicked ads as supervision signals. Then a student model is trained on both clicked and unclicked ads with knowledge distillation, performing uncertainty modeling to alleviate the inherent noise in pseudo-labels. Experiments on billion-scale datasets show that UKD outperforms previous debiasing methods. Online results verify that UKD achieves significant improvements.
We present a simple and effective pretraining strategy {D}en{o}ising {T}raining DoT for neural machine translation. Specifically, we update the model parameters with source- and target-side denoising tasks at the early stage and then tune the model normally. Notably, our approach does not increase any parameters or training steps, requiring the parallel data merely. Experiments show that DoT consistently improves the neural machine translation performance across 12 bilingual and 16 multilingual directions (data size ranges from 80K to 20M). In addition, we show that DoT can complement existing data manipulation strategies, i.e. curriculum learning, knowledge distillation, data diversification, bidirectional training, and back-translation. Encouragingly, we found that DoT outperforms costly pretrained model mBART in high-resource settings. Analyses show DoT is a novel in-domain cross-lingual pretraining strategy and could offer further improvements with task-relevant self-supervisions.
In this study, we aim to predict the plausible future action steps given an observation of the past and study the task of instructional activity anticipation. Unlike previous anticipation tasks that aim at action label prediction, our work targets at generating natural language outputs that provide interpretable and accurate descriptions of future action steps. It is a challenging task due to the lack of semantic information extracted from the instructional videos. To overcome this challenge, we propose a novel knowledge distillation framework to exploit the related external textual knowledge to assist the visual anticipation task. However, previous knowledge distillation techniques generally transfer information within the same modality. To bridge the gap between the visual and text modalities during the distillation process, we devise a novel cross-modal contrastive distillation (CCD) scheme, which facilitates knowledge distillation between teacher and student in heterogeneous modalities with the proposed cross-modal distillation loss. We evaluate our method on the Tasty Videos dataset. CCD improves the anticipation performance of the visual-alone student model by a large margin of 40.2% relatively in BLEU4. Our approach also outperforms the state-of-the-art approaches by a large margin.
Representation knowledge distillation aims at transferring rich information from one model to another. Common approaches for representation distillation mainly focus on the direct minimization of distance metrics between the models' embedding vectors. Such direct methods may be limited in transferring high-order dependencies embedded in the representation vectors, or in handling the capacity gap between the teacher and student models. Moreover, in standard knowledge distillation, the teacher is trained without awareness of the student's characteristics and capacity. In this paper, we explore two mechanisms for enhancing representation distillation using classifier sharing between the teacher and student. We first investigate a simple scheme where the teacher's classifier is connected to the student backbone, acting as an additional classification head. Then, we propose a student-aware mechanism that asks to tailor the teacher model to a student with limited capacity by training the teacher with a temporary student's head. We analyze and compare these two mechanisms and show their effectiveness on various datasets and tasks, including image classification, fine-grained classification, and face verification. In particular, we achieve state-of-the-art results for face verification on the IJB-C dataset for a MobileFaceNet model: TAR@(FAR=1e-5)=93.7\%. Code is available at https://github.com/Alibaba-MIIL/HeadSharingKD.
Deep neural networks are typically trained in a single shot for a specific task and data distribution, but in real world settings both the task and the domain of application can change. The problem becomes even more challenging in dense predictive tasks, such as semantic segmentation, and furthermore most approaches tackle the two problems separately. In this paper we introduce the novel task of coarse-to-fine learning of semantic segmentation architectures in presence of domain shift. We consider subsequent learning stages progressively refining the task at the semantic level; i.e., the finer set of semantic labels at each learning step is hierarchically derived from the coarser set of the previous step. We propose a new approach (CCDA) to tackle this scenario. First, we employ the maximum squares loss to align source and target domains and, at the same time, to balance the gradients between well-classified and harder samples. Second, we introduce a novel coarse-to-fine knowledge distillation constraint to transfer network capabilities acquired on a coarser set of labels to a set of finer labels. Finally, we design a coarse-to-fine weight initialization rule to spread the importance from each coarse class to the respective finer classes. To evaluate our approach, we design two benchmarks where source knowledge is extracted from the GTA5 dataset and it is transferred to either the Cityscapes or the IDD datasets, and we show how it outperforms the main competitors.
Feature regression is a simple way to distill large neural network models to smaller ones. We show that with simple changes to the network architecture, regression can outperform more complex state-of-the-art approaches for knowledge distillation from self-supervised models. Surprisingly, the addition of a multi-layer perceptron head to the CNN backbone is beneficial even if used only during distillation and discarded in the downstream task. Deeper non-linear projections can thus be used to accurately mimic the teacher without changing inference architecture and time. Moreover, we utilize independent projection heads to simultaneously distill multiple teacher networks. We also find that using the same weakly augmented image as input for both teacher and student networks aids distillation. Experiments on ImageNet dataset demonstrate the efficacy of the proposed changes in various self-supervised distillation settings.
In the SSLAD-Track 3B challenge on continual learning, we propose the method of COntinual Learning with Transformer (COLT). We find that transformers suffer less from catastrophic forgetting compared to convolutional neural network. The major principle of our method is to equip the transformer based feature extractor with old knowledge distillation and head expanding strategies to compete catastrophic forgetting. In this report, we first introduce the overall framework of continual learning for object detection. Then, we analyse the key elements' effect on withstanding catastrophic forgetting in our solution. Our method achieves 70.78 mAP on the SSLAD-Track 3B challenge test set.
Visible-infrared person re-identification (VI-ReID) has been challenging due to the existence of large discrepancies between visible and infrared modalities. Most pioneering approaches reduce intra-class variations and inter-modality discrepancies by learning modality-shared and ID-related features. However, an explicit modality-shared cue, i.e., body keypoints, has not been fully exploited in VI-ReID. Additionally, existing feature learning paradigms imposed constraints on either global features or partitioned feature stripes, which neglect the prediction consistency of global and part features. To address the above problems, we exploit Pose Estimation as an auxiliary learning task to assist the VI-ReID task in an end-to-end framework. By jointly training these two tasks in a mutually beneficial manner, our model learns higher quality modality-shared and ID-related features. On top of it, the learnings of global features and local features are seamlessly synchronized by Hierarchical Feature Constraint (HFC), where the former supervises the latter using the knowledge distillation strategy. Experimental results on two benchmark VI-ReID datasets show that the proposed method consistently improves state-of-the-art methods by significant margins. Specifically, our method achieves nearly 20$\%$ mAP improvements against the state-of-the-art method on the RegDB dataset. Our intriguing findings highlight the usage of auxiliary task learning in VI-ReID.
Advanced face swapping methods have achieved appealing results. However, most of these methods have many parameters and computations, which makes it challenging to apply them in real-time applications or deploy them on edge devices like mobile phones. In this work, we propose a lightweight Identity-aware Dynamic Network (IDN) for subject-agnostic face swapping by dynamically adjusting the model parameters according to the identity information. In particular, we design an efficient Identity Injection Module (IIM) by introducing two dynamic neural network techniques, including the weights prediction and weights modulation. Once the IDN is updated, it can be applied to swap faces given any target image or video. The presented IDN contains only 0.50M parameters and needs 0.33G FLOPs per frame, making it capable for real-time video face swapping on mobile phones. In addition, we introduce a knowledge distillation-based method for stable training, and a loss reweighting module is employed to obtain better synthesized results. Finally, our method achieves comparable results with the teacher models and other state-of-the-art methods.
Speech recognition on smart devices is challenging owing to the small memory footprint. Hence small size ASR models are desirable. With the use of popular transducer-based models, it has become possible to practically deploy streaming speech recognition models on small devices [1]. Recently, the two-pass model [2] combining RNN-T and LAS modules has shown exceptional performance for streaming on-device speech recognition. In this work, we propose a simple and effective approach to reduce the size of the two-pass model for memory-constrained devices. We employ a popular knowledge distillation approach in three stages using the Teacher-Student training technique. In the first stage, we use a trained RNN-T model as a teacher model and perform knowledge distillation to train the student RNN-T model. The second stage uses the shared encoder and trains a LAS rescorer for student model using the trained RNN-T+LAS teacher model. Finally, we perform deep-finetuning for the student model with a shared RNN-T encoder, RNN-T decoder, and LAS rescorer. Our experimental results on standard LibriSpeech dataset show that our system can achieve a high compression rate of 55% without significant degradation in the WER compared to the two-pass teacher model.
While existing federated learning approaches primarily focus on aggregating local models to construct a global model, in realistic settings, some clients may be reluctant to share their private models due to the inclusion of privacy-sensitive information. Knowledge distillation, which can extract model knowledge without accessing model parameters, is well-suited for this federated scenario. However, most distillation methods in federated learning (federated distillation) require a proxy dataset, which is difficult to obtain in the real world. Therefore, in this paper, we introduce a distributed three-player Generative Adversarial Network (GAN) to implement data-free mutual distillation and propose an effective method called FedDTG. We confirmed that the fake samples generated by GAN can make federated distillation more efficient and robust. Additionally, the distillation process between clients can deliver good individual client performance while simultaneously acquiring global knowledge and protecting data privacy. Our extensive experiments on benchmark vision datasets demonstrate that our method outperforms other federated distillation algorithms in terms of generalization.
We receive several essential updates on our smartphones in the form of SMS, documents, voice messages, etc. that get buried beneath the clutter of content. We often do not realize the key information without going through the full content. SMS notifications sometimes help by giving an idea of what the message is about, however, they merely offer a preview of the beginning content. One way to solve this is to have a single efficient model that can adapt and summarize data from varied sources. In this paper, we tackle this issue and for the first time, propose a novel Adaptive Beam Search to improve the quality of on-device abstractive summarization that can be applied to SMS, voice messages and can be extended to documents. To the best of our knowledge, this is the first on-device abstractive summarization pipeline to be proposed that can adapt to multiple data sources addressing privacy concerns of users as compared to the majority of existing summarization systems that send data to a server. We reduce the model size by 30.9% using knowledge distillation and show that this model with a 97.6% lesser memory footprint extracts the same or more key information as compared to BERT.