Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of ``small-scale inverted data for knowledge distillation". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation SSD-KD. In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., 10X less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD.
Spiking neural networks (SNNs) have attracted considerable attention for their event-driven, low-power characteristics and high biological interpretability. Inspired by knowledge distillation (KD), recent research has improved the performance of the SNN model with a pre-trained teacher model. However, additional teacher models require significant computational resources, and it is tedious to manually define the appropriate teacher network architecture. In this paper, we explore cost-effective self-distillation learning of SNNs to circumvent these concerns. Without an explicit defined teacher, the SNN generates pseudo-labels and learns consistency during training. On the one hand, we extend the timestep of the SNN during training to create an implicit temporal ``teacher" that guides the learning of the original ``student", i.e., the temporal self-distillation. On the other hand, we guide the output of the weak classifier at the intermediate stage by the final output of the SNN, i.e., the spatial self-distillation. Our temporal-spatial self-distillation (TSSD) learning method does not introduce any inference overhead and has excellent generalization ability. Extensive experiments on the static image datasets CIFAR10/100 and ImageNet as well as the neuromorphic datasets CIFAR10-DVS and DVS-Gesture validate the superior performance of the TSSD method. This paper presents a novel manner of fusing SNNs with KD, providing insights into high-performance SNN learning methods.
Optical and hybrid convolutional neural networks (CNNs) recently have become of increasing interest to achieve low-latency, low-power image classification and computer vision tasks. However, implementing optical nonlinearity is challenging, and omitting the nonlinear layers in a standard CNN comes at a significant reduction in accuracy. In this work, we use knowledge distillation to compress modified AlexNet to a single linear convolutional layer and an electronic backend (two fully connected layers). We obtain comparable performance to a purely electronic CNN with five convolutional layers and three fully connected layers. We implement the convolution optically via engineering the point spread function of an inverse-designed meta-optic. Using this hybrid approach, we estimate a reduction in multiply-accumulate operations from 17M in a conventional electronic modified AlexNet to only 86K in the hybrid compressed network enabled by the optical frontend. This constitutes over two orders of magnitude reduction in latency and power consumption. Furthermore, we experimentally demonstrate that the classification accuracy of the system exceeds 93% on the MNIST dataset.
Audio packet loss is an inevitable problem in real-time speech communication. A band-split packet loss concealment network (BS-PLCNet) targeting full-band signals was recently proposed. Although it performs superiorly in the ICASSP 2024 PLC Challenge, BS-PLCNet is a large model with high computational complexity of 8.95G FLOPS. This paper presents its updated version, BS-PLCNet 2, to reduce computational complexity and improve performance further. Specifically, to compensate for the missing future information, in the wide-band module, we design a dual-path encoder structure (with non-causal and causal path) and leverage an intra-model knowledge distillation strategy to distill the future information from the non-causal teacher to the casual student. Moreover, we introduce a lightweight post-processing module after packet loss restoration to recover speech distortions and remove residual noise in the audio signal. With only 40% of original parameters in BS-PLCNet, BS-PLCNet 2 brings 0.18 PLCMOS improvement on the ICASSP 2024 PLC challenge blind set, achieving state-of-the-art performance on this dataset.
Rolling bearing fault detection has developed rapidly in the field of fault diagnosis technology, and it occupies a very important position in this field. Deep learning-based bearing fault diagnosis models have achieved significant success. At the same time, with the continuous improvement of new signal processing technologies such as Fourier transform, wavelet transform and empirical mode decomposition, the fault diagnosis technology of rolling bearings has also been greatly developed, and it can be said that it has entered a new research stage. However, most of the existing methods are limited to varying degrees in the industrial field. The main ones are fast feature extraction and computational complexity. The key to this paper is to propose a lightweight bearing fault diagnosis model DKDL-Net to solve these challenges. The model is trained on the CWRU data set by decoupling knowledge distillation and low rank adaptive fine tuning. Specifically, we built and trained a teacher model based on a 6-layer neural network with 69,626 trainable parameters, and on this basis, using decoupling knowledge distillation (DKD) and Low-Rank adaptive (LoRA) fine-tuning, we trained the student sag model DKDL-Net, which has only 6838 parameters. Experiments show that DKDL-Net achieves 99.48% accuracy in computational complexity on the test set while maintaining model performance, which is 0.58% higher than the state-of-the-art (SOTA) model, and our model has lower parameters. Our code is available at Github link: https://github.com/SPBU-LiPengyi/DKDL-Net.git.
Large language models (LLMs) have achieved remarkable performance on Natural Language Processing (NLP) tasks, but they are hindered by high computational costs and memory requirements. Ternarization, an extreme form of quantization, offers a solution by reducing memory usage and enabling energy-efficient floating-point additions. However, applying ternarization to LLMs faces challenges stemming from outliers in both weights and activations. In this work, observing asymmetric outliers and non-zero means in weights, we introduce Dual Learnable Ternarization (DLT), which enables both scales and shifts to be learnable. We also propose Outlier-Friendly Feature Knowledge Distillation (OFF) to recover the information lost in extremely low-bit quantization. The proposed OFF can incorporate semantic information and is insensitive to outliers. At the core of OFF is maximizing the mutual information between features in ternarized and floating-point models using cosine similarity. Extensive experiments demonstrate that our TernaryLLM surpasses previous low-bit quantization methods on the standard text generation and zero-shot benchmarks for different LLM families. Specifically, for one of the most powerful open-source models, LLaMA-3, our approach (W1.58A16) outperforms the previous state-of-the-art method (W2A16) by 5.8 in terms of perplexity on C4 and by 8.2% in terms of average accuracy on zero-shot tasks.
Knowledge distillation (KD) is a widely adopted and effective method for compressing models in object detection tasks. Particularly, feature-based distillation methods have shown remarkable performance. Existing approaches often ignore the uncertainty in the teacher model's knowledge, which stems from data noise and imperfect training. This limits the student model's ability to learn latent knowledge, as it may overly rely on the teacher's imperfect guidance. In this paper, we propose a novel feature-based distillation paradigm with knowledge uncertainty for object detection, termed "Uncertainty Estimation-Discriminative Knowledge Extraction-Knowledge Transfer (UET)", which can seamlessly integrate with existing distillation methods. By leveraging the Monte Carlo dropout technique, we introduce knowledge uncertainty into the training process of the student model, facilitating deeper exploration of latent knowledge. Our method performs effectively during the KD process without requiring intricate structures or extensive computational resources. Extensive experiments validate the effectiveness of our proposed approach across various distillation strategies, detectors, and backbone architectures. Specifically, following our proposed paradigm, the existing FGD method achieves state-of-the-art (SoTA) performance, with ResNet50-based GFL achieving 44.1% mAP on the COCO dataset, surpassing the baselines by 3.9%.
We propose Hydra-MDP, a novel paradigm employing multiple teachers in a teacher-student model. This approach uses knowledge distillation from both human and rule-based teachers to train the student model, which features a multi-head decoder to learn diverse trajectory candidates tailored to various evaluation metrics. With the knowledge of rule-based teachers, Hydra-MDP learns how the environment influences the planning in an end-to-end manner instead of resorting to non-differentiable post-processing. This method achieves the $1^{st}$ place in the Navsim challenge, demonstrating significant improvements in generalization across diverse driving environments and conditions. More details by visiting \url{https://github.com/NVlabs/Hydra-MDP}.
Direct speech-to-speech translation (S2ST) has achieved impressive translation quality, but it often faces the challenge of slow decoding due to the considerable length of speech sequences. Recently, some research has turned to non-autoregressive (NAR) models to expedite decoding, yet the translation quality typically lags behind autoregressive (AR) models significantly. In this paper, we investigate the performance of CTC-based NAR models in S2ST, as these models have shown impressive results in machine translation. Experimental results demonstrate that by combining pretraining, knowledge distillation, and advanced NAR training techniques such as glancing training and non-monotonic latent alignments, CTC-based NAR models achieve translation quality comparable to the AR model, while preserving up to 26.81$\times$ decoding speedup.
Transformer-based retrieval and reranking models for text document search are often refined through knowledge distillation together with contrastive learning. A tight distribution matching between the teacher and student models can be hard as over-calibration may degrade training effectiveness when a teacher does not perform well. This paper contrastively reweights KL divergence terms to prioritize the alignment between a student and a teacher model for proper separation of positive and negative documents. This paper analyzes and evaluates the proposed loss function on the MS MARCO and BEIR datasets to demonstrate its effectiveness in improving the relevance of tested student models.
Policy Distillation (PD) has become an effective method to improve deep reinforcement learning tasks. The core idea of PD is to distill policy knowledge from a teacher agent to a student agent. However, the teacher-student framework requires a well-trained teacher model which is computationally expensive.In the light of online knowledge distillation, we study the knowledge transfer between different policies that can learn diverse knowledge from the same environment.In this work, we propose Online Policy Distillation (OPD) with Decision-Attention (DA), an online learning framework in which different policies operate in the same environment to learn different perspectives of the environment and transfer knowledge to each other to obtain better performance together. With the absence of a well-performance teacher policy, the group-derived targets play a key role in transferring group knowledge to each student policy. However, naive aggregation functions tend to cause student policies quickly homogenize. To address the challenge, we introduce the Decision-Attention module to the online policies distillation framework. The Decision-Attention module can generate a distinct set of weights for each policy to measure the importance of group members. We use the Atari platform for experiments with various reinforcement learning algorithms, including PPO and DQN. In different tasks, our method can perform better than an independent training policy on both PPO and DQN algorithms. This suggests that our OPD-DA can transfer knowledge between different policies well and help agents obtain more rewards.
There is increasing interest in distilling task-specific knowledge from large language models (LLM) to smaller student models. Nonetheless, LLM distillation presents a dual challenge: 1) there is a high cost associated with querying the teacher LLM, such as GPT-4, for gathering an ample number of demonstrations; 2) the teacher LLM might provide imperfect outputs with a negative impact on the student's learning process. To enhance sample efficiency within resource-constrained, imperfect teacher scenarios, we propose a three-component framework leveraging three signal types. The first signal is the student's self-consistency (consistency of student multiple outputs), which is a proxy of the student's confidence. Specifically, we introduce a ``teaching assistant'' (TA) model to assess the uncertainty of both the student's and the teacher's outputs via confidence scoring, which serves as another two signals for student training. Furthermore, we propose a two-stage training schema to first warm up the student with a small proportion of data to better utilize student's signal. Experiments have shown the superiority of our proposed framework for four complex reasoning tasks. On average, our proposed two-stage framework brings a relative improvement of up to 20.79% compared to fine-tuning without any signals across datasets.
Arabic is known to present unique challenges for Automatic Speech Recognition (ASR). On one hand, its rich linguistic diversity and wide range of dialects complicate the development of robust, inclusive models. On the other, current multilingual ASR models are compute-intensive and lack proper comprehensive evaluations. In light of these challenges, we distill knowledge from large teacher models into smaller student variants that are more efficient. We also introduce a novel human-annotated dataset covering five under-represented Arabic dialects for evaluation. We further evaluate both our models and existing SoTA multilingual models on both standard available benchmarks and our new dialectal data. Our best-distilled model's overall performance ($45.0$\% WER) surpasses that of a SoTA model twice its size (SeamlessM4T-large-v2, WER=$47.0$\%) and its teacher model (Whisper-large-v2, WER=$55.1$\%), and its average performance on our new dialectal data ($56.9$\% WER) outperforms all other models. To gain more insight into the poor performance of these models on dialectal data, we conduct an error analysis and report the main types of errors the different models tend to make. The GitHub repository for the project is available at \url{https://github.com/UBC-NLP/distill-whisper-ar}.
Existing Incremental Object Detection (IOD) methods partially alleviate catastrophic forgetting when incrementally detecting new objects in real-world scenarios. However, many of these methods rely on the assumption that unlabeled old-class objects may co-occur with labeled new-class objects in the incremental data. When unlabeled old-class objects are absent, the performance of existing methods tends to degrade. The absence can be mitigated by generating old-class samples, but it incurs high costs. This paper argues that previous generation-based IOD suffers from redundancy, both in the use of generative models, which require additional training and storage, and in the overproduction of generated samples, many of which do not contribute significantly to performance improvements. To eliminate the redundancy, we propose Inversed Objects Replay (IOR). Specifically, we generate old-class samples by inversing the original detectors, thus eliminating the necessity of training and storing additional generative models. We propose augmented replay to reuse the objects in generated samples, reducing redundant generations. Moreover, we propose high-value knowledge distillation focusing on the positions of old-class objects overwhelmed by the background, which transfers the knowledge to the incremental detector. Extensive experiments conducted on MS COCO 2017 demonstrate that our method can efficiently improve detection performance in IOD scenarios with the absence of old-class objects.
Data often arrives in sequence over time in real-world deep learning applications such as autonomous driving. When new training data is available, training the model from scratch undermines the benefit of leveraging the learned knowledge, leading to significant training costs. Warm-starting from a previously trained checkpoint is the most intuitive way to retain knowledge and advance learning. However, existing literature suggests that this warm-starting degrades generalization. In this paper, we advocate for warm-starting but stepping out of the previous converging point, thus allowing a better adaptation to new data without compromising previous knowledge. We propose Knowledge Consolidation and Acquisition (CKCA), a continuous model improvement algorithm with two novel components. First, a novel feature regularization (FeatReg) to retain and refine knowledge from existing checkpoints; Second, we propose adaptive knowledge distillation (AdaKD), a novel approach to forget mitigation and knowledge transfer. We tested our method on ImageNet using multiple splits of the training data. Our approach achieves up to $8.39\%$ higher top1 accuracy than the vanilla warm-starting and consistently outperforms the prior art with a large margin.
Lensless cameras, innovatively replacing traditional lenses for ultra-thin, flat optics, encode light directly onto sensors, producing images that are not immediately recognizable. This compact, lightweight, and cost-effective imaging solution offers inherent privacy advantages, making it attractive for privacy-sensitive applications like face verification. Typical lensless face verification adopts a two-stage process of reconstruction followed by verification, incurring privacy risks from reconstructed faces and high computational costs. This paper presents an end-to-end optimization approach for privacy-preserving face verification directly on encoded lensless captures, ensuring that the entire software pipeline remains encoded with no visible faces as intermediate results. To achieve this, we propose several techniques to address unique challenges from the lensless setup which precludes traditional face detection and alignment. Specifically, we propose a face center alignment scheme, an augmentation curriculum to build robustness against variations, and a knowledge distillation method to smooth optimization and enhance performance. Evaluations under both simulation and real environment demonstrate our method outperforms two-stage lensless verification while enhancing privacy and efficiency. Project website: \url{lenslessface.github.io}.
Self-supervised learning (SSL) is increasingly attractive for pre-training encoders without requiring labeled data. Downstream tasks built on top of those pre-trained encoders can achieve nearly state-of-the-art performance. The pre-trained encoders by SSL, however, are vulnerable to backdoor attacks as demonstrated by existing studies. Numerous backdoor mitigation techniques are designed for downstream task models. However, their effectiveness is impaired and limited when adapted to pre-trained encoders, due to the lack of label information when pre-training. To address backdoor attacks against pre-trained encoders, in this paper, we innovatively propose a mutual information guided backdoor mitigation technique, named MIMIC. MIMIC treats the potentially backdoored encoder as the teacher net and employs knowledge distillation to distill a clean student encoder from the teacher net. Different from existing knowledge distillation approaches, MIMIC initializes the student with random weights, inheriting no backdoors from teacher nets. Then MIMIC leverages mutual information between each layer and extracted features to locate where benign knowledge lies in the teacher net, with which distillation is deployed to clone clean features from teacher to student. We craft the distillation loss with two aspects, including clone loss and attention loss, aiming to mitigate backdoors and maintain encoder performance at the same time. Our evaluation conducted on two backdoor attacks in SSL demonstrates that MIMIC can significantly reduce the attack success rate by only utilizing <5% of clean data, surpassing seven state-of-the-art backdoor mitigation techniques.
Benefiting from well-trained deep neural networks (DNNs), model compression have captured special attention for computing resource limited equipment, especially edge devices. Knowledge distillation (KD) is one of the widely used compression techniques for edge deployment, by obtaining a lightweight student model from a well-trained teacher model released on public platforms. However, it has been empirically noticed that the backdoor in the teacher model will be transferred to the student model during the process of KD. Although numerous KD methods have been proposed, most of them focus on the distillation of a high-performing student model without robustness consideration. Besides, some research adopts KD techniques as effective backdoor mitigation tools, but they fail to perform model compression at the same time. Consequently, it is still an open problem to well achieve two objectives of robust KD, i.e., student model's performance and backdoor mitigation. To address these issues, we propose RobustKD, a robust knowledge distillation that compresses the model while mitigating backdoor based on feature variance. Specifically, RobustKD distinguishes the previous works in three key aspects: (1) effectiveness: by distilling the feature map of the teacher model after detoxification, the main task performance of the student model is comparable to that of the teacher model; (2) robustness: by reducing the characteristic variance between the teacher model and the student model, it mitigates the backdoor of the student model under backdoored teacher model scenario; (3) generic: RobustKD still has good performance in the face of multiple data models (e.g., WRN 28-4, Pyramid-200) and diverse DNNs (e.g., ResNet50, MobileNet).
Knowledge distillation (KD) has been shown to be highly effective in guiding a student model with a larger teacher model and achieving practical benefits in improving the computational and memory efficiency for large language models (LLMs). State-of-the-art KD methods for LLMs mostly rely on minimizing explicit distribution distance between teacher and student probability predictions. Instead of optimizing these mandatory behaviour cloning objectives, we explore an imitation learning strategy for KD of LLMs. In particular, we minimize the imitation gap by matching the action-value moments of the teacher's behavior from both on- and off-policy perspectives. To achieve this action-value moment-matching goal, we propose an adversarial training algorithm to jointly estimate the moment-matching distance and optimize the student policy to minimize it. Results from both task-agnostic instruction-following experiments and task-specific experiments demonstrate the effectiveness of our method and achieve new state-of-the-art performance.
This thesis aims to investigate the feasibility of knowledge transfer between neural networks for medical image segmentation tasks, specifically focusing on the transfer from a larger multi-task "Teacher" network to a smaller "Student" network. In the context of medical imaging, where the data volumes are often limited, leveraging knowledge from a larger pre-trained network could be useful. The primary objective is to enhance the performance of a smaller student model by incorporating knowledge representations acquired by a teacher model that adopts a multi-task pre-trained architecture trained on CT images, to a more resource-efficient student network, which can essentially be a smaller version of the same, trained on a mere 50% of the data than that of the teacher model. To facilitate knowledge transfer between the two models, we devised an architecture incorporating multi-scale feature distillation and supervised contrastive learning. Our study aims to improve the student model's performance by integrating knowledge representations from the teacher model. We investigate whether this approach is particularly effective in scenarios with limited computational resources and limited training data availability. To assess the impact of multi-scale feature distillation, we conducted extensive experiments. We also conducted a detailed ablation study to determine whether it is essential to distil knowledge at various scales, including low-level features from encoder layers, for effective knowledge transfer. In addition, we examine different losses in the knowledge distillation process to gain insights into their effects on overall performance.