• What we'll do
In this meetup being sponsored by AMD, we'll have two great speakers. Mike Schmit from AMD will talk about how to make your AI application better with a proper dataset and the common problems from training with bad data. Mike will show numerous examples of how errors in your dataset can lead to bad results.
Waleed Abdulla will talk about modern object detection and instance segmentation network. Tentative agenda is:
- Pizza and networking 6-7PM
- Welcome and sponsor message 7PM
- 7-745: First presentation and Q&A
-[masked]: Second presentation and Q&A
Presentation 1: How to Create the Right Dataset for Your Application
When training deep neural networks, having the right training data is key. In this talk, I will explore what makes a successful training dataset, common pitfalls and how to set realistic expectations. I’ll illustrate 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. My results are based on my personal dataset of over 10,000 hand-labeled images from around the world.
Speaker bio: Mike Schmit is the Director of Software Engineering for computer vision and machine learning at AMD. Mike has been immersed in code optimization for many years. He was the chief software architect for the first 8086 to control experiments in the Space Shuttle, authored the formative book on optimizing code for the Pentium and developed and managed the team that built the first software DVD player. Shortly after that he joined ATI and managed the software video codec team, for many years, which eventually began working on computer vision optimizations and then the OpenVX computer vision standard. Mike has given many industry talks on his team’s optimizations for massively parallel GPUs including recent talks on 360 Video stitching at VRLA, SVVR and Oculus OC3.
Presentation 2: Learn how Modern Object Detection and Instance Segmentation Networks Work
In this presentation, Waleed will explain how object detection and instance segmentation models work, and will cover lessons learned from his experience building such model that got picked up and used by thousands of deep learning developers.
Speaker bio: Waleed Abdulla is a deep learning engineer focusing on computer vision applications. He writes about deep learning and builds open source projects. His most recent project, Mask RCNN, is one of the top instance segmentation tools on github, used by thousands of deep learning developers. He's an independent consultant, while also working on his next startup project. Before getting into deep learning, he built a startup in the news and social media space, raised VC funding, and was in 500Startups and the Facebook fbFund before that. He also served on the board of Hacker Dojo, a non-profit co-working space. And he's often active in organizing technology events, meetups, and hackathons.