Abstract

In this interactive workshop, we will collaboratively build a robust object segmentation AI, trained a common item, and later identify its flaws and limitations. Participants will be given access to an interactive dataset platform where they can view, label, and capture additional image data to train a neural network. We will train a model during the lecture, and run it in real-time to identify its strengths and weaknesses and how this may affect real-world applications.

Deep learning allows computer vision systems to skyrocket form a number of hand-crafted heuristics in traditional vision engineering, to millions of automatically learned parameters, allowing it to learn almost anything. Backed by significant hype, it seems to break through all the challenges presented by AI, but where does it still fail, and why?

Topics covered include: A brief history of computer vision, what data can and cannot be learned, strengths and limitations of supervised learning, training a Mask-RCNN model via Pytorch, monitoring an AI’s performance.

Free pizza and beer available after the talk!