Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks

By Lin Wang, Kuk-Jin Yoon. 2020

965 citations recorded by the source corpus.

Deep neural models, in recent years, have been successful in almost every field, even solving the most complex problem statements. However, these models are huge in size with millions (and even billions) of parameters, demanding heavy computation power and failing to be deployed on edge devices. Besides, the performance boost is highly dependent on redundant labeled data. To achieve faster speeds and to handle the problems caused by the lack of labeled data, knowledge distillation (KD) has been proposed to transfer information learned from one model to another. KD is often characterized by the so-called ‘Student-Teacher’ (S-T) learning framework and has been broadly applied in model compression and knowledge transfer. This paper is about KD and S-T learning, which are being actively studied in recent years. First, we aim to provide explanations of what KD is and how/why it works. Then, we provide a comprehensive survey on the recent progress of KD methods together with S-T frameworks typically used for vision tasks. In general, we investigate some fundamental questions that have been driving this research area and thoroughly generalize the research progress and technical details. Additionally, we systematically analyze the research status of KD in vision applications. Finally, we discuss the potentials and open challenges of existing methods and prospect the future directions of KD and S-T learning.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Efficient Acceleration of Deep Learning Inference on Resource-Constrained Edge Devices: A Review (2023, methodology): “KD is a compression method that mimics the behavior of a large DL model into a lightweight model by transferring learned knowledge [178].”
A Full Dive Into Realizing the Edge-Enabled Metaverse: Visions, Enabling Technologies, and Challenges (2022, methodology): “A visualization of the student-teacher network paradigm [217].”
Learning Compatible Embeddings (2021, background): “There exists a great amount of knowledge sources and knowledge types in this field (see [13, 48] for recent reviews).”
LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS (2023, background): “Knowledge Distillation Knowledge Distillation (KD) is widely adopted in various machine learning tasks [50, 53, 60, 61].”
Large-Scale Generative Data-Free Distillation (2020, methodology): “. This limits our ability to extend and modify existing neural networks. Recently knowledge distillation [23,46] has emerged as a robust way of transferring knowledge across different neural networks [6,19,52]. In the teacher*Work done as part of the Google AI Residency. student framework, the stude”
Decoupled Knowledge Distillation (2022, background): “Previous works [3, 36] explained this phenomenon with the large capacity gap between big teachers and small students.”
Balanced Knowledge Distillation for Long-tailed Learning (2021, background): “Orthogonally to the previous works, we disentangle the contradiction between the two goals from the perspective of teacher-student learning [16, 52, 49].”
Student-Teacher Feature Pyramid Matching for Anomaly Detection (2021, background): “For the former, since knowledge is distilled from a ResNet-18 [16] into a lightweight teacher network, the big gap between their model capacities [42] tends to incur loss of important information.”
A Fast Knowledge Distillation Framework for Visual Recognition (2021, background): “For the broader overviews of related methods for knowledge distillation, please refer to [12,43].”
Compression of Deep Learning Models for Text: A Survey (2020, background): “While there are other surveys in the broad area of model compression [14, 22] also specifically on knowledge distillation [126], they are either old or focus on computer vision related problems.”
Looking Back on Learned Experiences For Class/task Incremental Learning (2022, methodology): “For the distillation purpose, we adopt knowledge distillation from network output (logits), known as dark knowledge Hinton et al. (2015); Wang & Yoon (2020); Buzzega et al. (2020), and our objective is to minimize the Euclidean distance between the stored logits in the transfer set O = [ o 1 ( x )…”