Avoiding Pitfalls When Building Deep Learning Vision Systems

Deep learning is the future of visual inspection. To build an accurate and robust deep learning system, teams traditionally focus on improving either the model or the algorithm. However, this approach has proven to be inadequate in a production setting. A deep learning system usually fails to meet the requirements of a production environment when the data used to train the system has not been correctly sorted and labeled, not because of the model used. This explains why many AI teams typically spend 80% of their time on data preparation and only 20% on model training. Based on the real-world experience of building and shipping deep learning-based solutions for industry leaders such as Stanley Black & Decker, Bibireata sheds light on common pitfalls in data preparation and how to avoid them.

About the presenter
Daniel Bibireata is vice president of engineering at Landing AI, a company that offers an end-to-end AI platform that enables customers to build, deploy, and manage deep learning-based visual inspection solutions. Prior to working at Landing AI, Daniel was a principal engineer at Amazon for 15 years, where he worked on the computer vision technology behind the Amazon Go stores.