庖丁解牛与盲人摸象.pdfVIP

VALSE 2018, 大连, 2018.4.22 深度学习实践： 庖丁解牛与盲人摸象 吴建鑫 南京大学教授 MINIEYE 首席科学家 1 What is my interest in deep learning? 2 Co-occurrence is not enough （彬彬）： （妈妈每天晚上）： 狗狗猫猫！ 彬彬好好睡觉哦， 睡着以后会有狗狗 猫猫陪彬彬玩的哦 （爸爸中午）：现 在给你脱衣服，过 会儿彬彬该干嘛呢？ 感谢彬彬小朋友提供此例 3 DL engineering- research heavy, too! 比亚迪的ADAS 前装系统为MINIEYE提供 4 Understanding Accelerating ■ What is in the representation? ■ How to accelerate the inference? ■ How is the representation generated? ■ What is the redundancy? – Can we improve or customize it? 5 盲人摸象：利用已有DL模型或方法，即便我们不理解该表示，我们是否能做些什么？ 6 SCDA: What is object? Object vs. non-object (background) Selective Convolutional Descriptor Aggregation for Fine-Grained Image Retrieval Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, Zhi-Hua Zhou IEEE Transactions on Image Processing, 2017, 26(6): 2868-2881 7 Distributed representation  Knowledge Visual attributes