Ars Electronica 2019 - AI Lab

Transcript

1. AI Lab Melina Undesser (AT) Sebastian Lindinger (AT)
2. Computers are dumb! They can only do what we program them to do.
3. They can only work with zeroes and ones. The world does not consist of just zeroes and ones.
4. Computers will never learn to see, hear, feel, smell… Or will they?
5. Two young people, no computer science professionals Mentors from CoderDojo Linz Discover AI for face detection Combine face detection with game development Can they teach the computer to detect faces and emotions? Can they build games controlled by head movement and facial expressions?
6. Just like humans, computers learn from experience. So, you need a lot of training data.
7. For training a good face detection AI, you need thousands of training images Diversity is super important to avoid a biased AI Data is the new gold Many universities make their image databases available for research purposes (e.g. XM2VTS) Commercial companies try to gather data through e.g. free services For startup, it is hard to compete with internet giants Snowball effect A lot of power controlled by only a few organizations
8. Manual annotation and labelling done by humans
9. Muffin or Chihuahua? Image Source: Mariya Yao
10. We all help for free (e.g. reCAPTCHA)
11. Mechanical Turk Fake chess-playing machine constructed in the late 18th century Amazon Mechanical Turk Crowdsourcing website People perform discrete on-demand tasks that computers are currently unable to do „Turkers“ People labelling AI for large enterprises Blue-collar job of the AI era
12. The result of AI training is a model
13. Video capturing in Browser Face detection: face-api.js Completely done in Browser → privacy Uses Tensorflow.js, pretrained models Detection sent back to server Websockets Separate detection and processing Scalability
14. Scratch Extension Based on React Receives detection from server via Websockets Control game with your face Face movement Expression (happiness)
16. Try our face detection Car Race Don‘t Laugh Challenge Emotional Pong Discuss AI with us How did we build the project? What did we learn? What can you build? Develop your own face-controlled Scratch game