Timezone: »

 
Poster
MentorNet: Learning Data-Driven Curriculum for Very Deep Neural Networks on Corrupted Labels
Lu Jiang · Zhengyuan Zhou · Thomas Leung · Li-Jia Li · Li Fei-Fei

Thu Jul 12 09:15 AM -- 12:00 PM (PDT) @ Hall B #113

Recent deep networks are capable of memorizing the entire data even when the labels are completely random. To overcome the overfitting on corrupted labels, we propose a novel technique of learning another neural network, called MentorNet, to supervise the training of the base deep networks, namely, StudentNet. During training, MentorNet provides a curriculum (sample weighting scheme) for StudentNet to focus on the sample the label of which is probably correct. Unlike the existing curriculum that is usually predefined by human experts, MentorNet learns a data-driven curriculum dynamically with StudentNet. Experimental results demonstrate that our approach can significantly improve the generalization performance of deep networks trained on corrupted training data. Notably, to the best of our knowledge, we achieve the best-published result on WebVision, a large benchmark containing 2.2 million images of real-world noisy labels.

Author Information

Lu Jiang (Google)

Lu Jiang is a senior research scientist at Google Research and an adjunct faculty at Carnegie Mellon University. He received a Ph.D. in Artificial Intelligence (Language Technology) from Carnegie Mellon University in 2017. Lu's primary interests lie in the interdisciplinary field of Multimedia, Machine Learning, and Computer Vision. He is the recipient of Yahoo Fellow and Erasmus Mundus Scholar. He received Best paper nomination at ACL and ICMR; Best poster at IEEE SLT; Best system in NIST TRECVID. He was a key contributor to IARPA Aladdin project, and helped create YouTube-8M dataset and AutoML in Google. He served as an area chair for ACM Multimedia, AVSS; the PC for premier conferences such as CVPR, ICML, ICCV, AAAI and IJCAI; the reviewer for NSF SBIR/STTR Panel, Google Faculty Award, and Google Research Hiring Committee; the journal reviewer for JMLR, TPAMI, JAIR, TMM, CVIU, etc.

Zhengyuan Zhou (Stanford University)
Thomas Leung (Google Inc)
Li-Jia Li (Google)
Li Fei-Fei (Stanford University & Google)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors