[Lecture] Building a Modern Machine Learning Accelerator - Day 1

Part 1: Some History, Hardware Lottery

---

AI Winter for Neural Networks

Marvin Minsky wrote in 1969:

• “Neural networks” cannot learn the XOR relationship
• “Neural networks” only learn linearly separable problems
• “Neural networks” become too large for complex knowledge

But, “Single layer” neural networks had these problems.
→ “Deep” neural networks (DNN) are better.

but DNN is still computationally intractable to train.

Hardware Lottery Winners: Genernal-Purpose CPU Threads