Building intelligent robots
0 likes25 views
This document discusses computer vision and ethics in artificial intelligence. It covers topics like deep learning for computer vision, multi-task learning using uncertainty, and video semantic segmentation. It also discusses the trolley problem in ethics and provides examples of unethical AI today. Some concrete problems in autonomous vehicle safety are identified, including trust, fairness, interpretability and avoiding reward hacking. The document concludes that while deep learning can help machines perceive and act from data, models also need to understand ethics themselves to be applied safely in real-world situations.
Building intelligent robots
- 8. Deep Learning for Computer Vision • Powerful framework for understanding high dimensional data like images, videos, speech, text • With enough training data, they outperform human baselines for recognition tasks • Typically computer vision models contain ~10 million parameters, take 3+ days to train on a GPU
- 13. 1. The Trolley Problem, Wikipedia. https://en.wikipedia.org/wiki/Trolley_problem 2. Rowan McAllister, Yarin Gal, Alex Kendall, Mark van der Wilk, Amar Shah, Roberto Cipolla, and Adrian Weller. Concrete Problems for Autonomous Vehicle Safety: Advantages of Bayesian Deep Learning. IJCAI, 2017. The Trolley Problem Do you… 1. Do nothing, and the trolley kills the five people on the main track. 2. Pull the lever, diverting the trolley onto the side track where it will kill one person. Which is the most ethical choice?
- 14. Examples of Un-Ethical AI, Today • US Justice System Re-Offending Rate algorithm.. Biased against minorities https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing • Google’s speech recognition system is better at male voices https://makingnoiseandhearingthings.com/2016/07/12/googles-speech-recognition-has-a-gender-bias/ • Many self-driving cars only work in California https://www.ft.com/content/4377b4c0-0479-11e7-aa5b-6bb07f5c8e12
- 15. Concrete Problems for AI Safety • Trust • Improve model performance and accuracy • Models understanding uncertainty in decisions • Fairness • Improve data efficiency and reduce bias • Algorithms which require less training data and generalise better • Honesty • Interpretability of results and model saliency • Causal reasoning • Avoid reward hacking 1. Rowan McAllister, Yarin Gal, Alex Kendall, Mark van der Wilk, Amar Shah, Roberto Cipolla, and Adrian Weller. Concrete Problems for Autonomous Vehicle Safety: Advantages of Bayesian Deep Learning. IJCAI, 2017.
- 17. Conclusions • We need machine learning to scale to hard problems with intelligent robots – hand engineering isn’t good enough • We can learning to perceive and act from data with deep learning • We cannot explicitly reason about ethical situations on a case-by-case basis – our models need to understand ethics themselves • Computer vision is holding back robotics from real-world applications
- 18. alexgkendall.com/publica8ons/ @alexgkendall agk34@cam.ac.uk Thank you to the amazing people who made this work possible: Roberto Cipolla, Vijay Badrinarayanan, Yarin Gal, Tom Roddick, MaMhew Grimes, Adrian Weller, Amar Shah, Hayk Mar8rosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abe Bachrach, Adam Bry