Institute For The Future

Futurist Book Reviews

RECOMMENDED READING
The Alignment Problem: Machine Learning and Human Values

In 2017, a group of Stanford computers scientists and physicians published a widely read study showing that convolutional neural networks were as good as trained dermatologists at diagnosing a wide variety of skin cancers--a finding that hinted at the potential to deploy high quality, essentially free cancer diagnostics inside the phones of the near future. A year later, the study's authors had to amend their findings. While the algorithm was quite good at diagnosing early stage skin cancers, it frequently misdiagnosed any image of a mole that happened to have a ruler in it as cancerous because physicians typically photograph suspicious moles next to a ruler.

This example of the potential and pitfalls of machine learning systems is one of myriad fascinating examples in Brian Christian's latest book, The Alignment Problem: Machine Learning and Human Values, which explores the origins, present state and potential future of AI. Jumping seamlessly between memorable stories and academic theory, The Alignment Problem is not just fun to read but provides a level of technical explanation about how machine learning models work--and how they go astray--that is rare to find in popular books. Rather than hype or easy answers, The Alignment Problem provides a deep understanding of the kinds of dilemmas and questions we'll face as we increasingly entrust more and more of our lives to machines in the decades to come.

From Bradley Kreit, Director, IFTF Vantage Research