An AI Primer with Wojciech Zaremba

TL;DR

Wojciech Zaremba explains AI, machine learning, and deep learning, detailing how deep neural networks with many steps of computation outperform shallow models. He describes OpenAI’s projects in robotics (manipulation) and playing many computer games, noting that supervised learning is the only paradigm ready for business while reinforcement learning faces fundamental issues.

Key insights

• AI is extremely broad; early Google Search (PageRank) was AI without machine learning, using only hand-coded rules.
• Deep learning is simply using multiple steps of computation on data, as opposed to single-step models that were mathematically easier but weaker.
• The key enabler for deep neural networks was realizing that proper weight initialization (controlling magnitude so outputs don’t blow up or shrink) allows stochastic gradient descent to work, a fact t