Recent Advances in Efficient Deep Learning for Computer Vision
Tutorial at Asian Conference on Computer Vision 2022
Deep learning has revolutionized artificial intelligence and is widely recognised as the technology leading the fourth industrial revolution. In general, the tremendous success of deep learning is attributed to its effective use of both the vast computational resources now available and large amounts of labeled data. In spite of the huge excitement generated by the recent developments, deep learning models have become formidably large and computationally intensive, hindering the massive deployment of these models to resource-limited edge platforms (e.g., robot, smart phone, autonomous vehicle) at low-latency and low-power for real-world AI applications. To tackle the efficiency bottlenecks of deep learning, a dominant direction of recent studies in computer vision is to develop energy-efficient, real-time and high-performance compact models, where the techniques include but not limit to low-bit quantization, model pruning, neural architecture search, efficient hardware and software codesign, light-weight module design, dynamic neural networks, and theorem for model compression, etc. The aim of this tutorial is to present the audience with recent advances of efficient deep learning from both theoretical and application perspectives, as well as to discuss state-of-the-art literature, existing challenges, and motivate future research that will prompt disruptive progress in the emerging field of efficient computer vision, benefitting not only researchers and industry stakeholders but also the environment for green AI.
The tutorial page can be found at