The Frontiers of Machine Learning and AI
summary
Deep learning is framed as a rebranding of 1970s neural networks but remains data-hungry, compute-intensive and poor at representing uncertainty, famously misclassifying images with tiny pixel noise into an adversarial example (school bus → ostrich). The proposed fix is probabilistic machine learning using Bayes rule (dating to the 18th century) that averages over all classifiers via priors and posteriors; that method yields better uncertainty estimates and motivates recent Bayesian deep learning workshops at NeurIPS that show the approach scaling and promising results.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:05Limitations of Deep Learning
- 1:17Adversarial Examples in Deep Learning
- 1:55Reverend Thomas Bayes FRS (1702-1761)
- 3:18Probabilistic Machine Learning
- 3:29Bayesian Machine Learning How does it work?
- 4:09Bayesian Deep Learning
related talks
46mKnow What You Don't Know: Tools to Understand Uncertainty in DL
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
01h 05mNavigating the Security and Privacy Landscape of Modern AI
35mUncertainty in deep learning
29mDeep Learning, Neuroscience and the future of AI
45mIntroduction To Deep Learning
32mStructured Deep Learning with Probabilistic Neural Prog
03mRethinking software engineering in the AI era