MIC 2018 - Malong Technical Talk

Bio:
Matt Scott is Co-Founder and CTO of Malong Technologies, an award-winning artificial intelligence startup based in China. The company produces leading computer vision technologies for the retail and medical industries. Matt has 15+ years R&D experience in computer vision and machine learning.

He is a Senior Member of the IEEE (SMIEEE), has published 70+ patents, and over a dozen research papers in top scientific conferences and journals, including cover featured articles in the Proceedings of the IEEE and IEEE Computer. At CVPR 2017, Matt and his team won first place in the WebVision Challenge, a worldwide computer vision contest from Google.
In 2014, Matt co-founded Malong Technologies with CEO Dinglong Huang. The company now has 150+ employees and recently completed a B-Round of venture capital funding led by Softbank in China. Accenture also chose Malong as its first investment in China, as part of a strategic alliance announced in July 2018. Matt leads the R&D efforts of the company creating state-of-the-art computer vision technology and products. Malong was selected as a 2018 Technology Pioneer by the World Economic Forum.

Prior to Malong, he was at Microsoft for 10 years, working as a senior research development lead at Microsoft Research on computer vision, machine learning, and NLP. In 2018, Matt was recognized by Fast Company as one of the 100 Most Creative People in Business in China. His work has been featured in The Wall Street Journal, Forbes, The New York Times, the Financial Times, PBS Nightly News, CNBC, and The Discovery Channel, among other media.
Matt is a proud alumnus of Boston University.

Title:
CurriculumNet: Weakly Supervised Learning from Large-Scale Web Images

Abstract:
We present a simple yet efficient approach capable of training deep neural networks on large-scale weakly-supervised web images, which are crawled raw from the Internet by using text queries, without any human annotation. We develop a principled learning strategy by leveraging curriculum learning, with the goal of handling a massive amount of noisy labels and data imbalance effectively. We design a new learning curriculum by measuring the complexity of data using its distribution density in a feature space, and rank the complexity in an unsupervised manner. This allows for an efficient implementation of curriculum learning on large-scale web images, resulting in a high-performance CNN model, where the negative impact of noisy labels is reduced substantially. Importantly, we show by experiments that those images with highly noisy labels can surprisingly improve the generalization capability of the model, by serving as a manner of regularization. Our approaches obtain state-of-the-art performance on four benchmarks: WebVision, ImageNet, Clothing-1M and Food-101. With an ensemble of multiple models, we achieved a top-5 error rate of 5.2% on the WebVision challenge for 1000-category classification. This result was the top performance by a wide margin, outperforming second place by a nearly 50% relative error rate. Code and models are available at this URL: https://github.com/MalongTech/researc....