Mike Schmit, Director of Software Engineering for computer vision and machine learning at AMD, presents the “Bad Data, Bad Network, or: How to Create the Right Dataset for Your Application” tutorial at the May 2018 Embedded Vision Summit.
When training deep neural networks, having the right training data is key. In this talk, Schmit explores what makes a successful training dataset, common pitfalls and how to set realistic expectations. He illustrates these ideas using several object classification models that have won the annual ImageNet challenge. By analyzing accurate and inaccurate classification examples (some humorous and some amazingly accurate) you will gain intuition on the workings of neural networks. Schmit’s results are based on his personal dataset of over 10,000 hand-labeled images from around the world.