The main idea of this book is to demonstrate in greatly simplified form how AI works, how statistical methods used, illustrate it with multiple entertaining examples, and assure people that AI is coming, but it is not scary and will work together with humans even if it sometimes will control many functions controlled currently by humans such as driving, robotic surgeries, and so on.
Here author discusses AI and its popularity. Author defines it as trainable algorithm when programmer does not define how process goes, but rather how program trains itself by processing multitude of data and the rules of probability. Then he goes through a brief history of AI and discusses anxieties it created due to resent dramatic improvement in its performance.
- The Refugee
The chapters starts with discussion of Netflix, the company that excel in using probabilities derived from huge customer base to propose movie selection. It is using conditional probability to do it. The author retells the story of Abraham Wald – refugee from Hungary who developed statistical analysis methods to generate recommendation on survivability and consequently protection of different parts of a bomber plane, and improve quality inspection protocols to decrease production defects. His work with planes included adjustments for survival bias, by using same assumption for all types of damage and calculating survival probability. After that author demonstrates application of Wald algorithm to Netflix processing.
- The Candlestick Maker
This is about pattern recognition. It starts with the funny story of thieves stealing toilet paper rolls in Beijing, which led to implementation of face recognition in public toilets and draconian requirements it caused for people to present their face in readable by computer form. The author goes a bit into details of pattern recognition based on statistical evaluation of inputs and outputs. After that author moves to astronomy and the story of Henrietta Leavitt who developed method for calculating distance to starts by using pulsation and brightness. One of results was reevaluation of distance to various starts and Hubble’s discovery of Andromeda being far outside of Milky Way. From here author formulates key ideas of pattern recognition in AI:
- In AI, a “pattern” is a prediction rule that maps an input to an output.
- “Learning a pattern” means fitting a good prediction rule to data set.
Here is a graphic example:
Author makes a very important point that in image recognition AI now can outperform humans. For example image “Alaskan Malamute” produced 5% error rate for humans, but only 3% for AI in 2016, down from 25% in 2011. This result was produced by 22-layer neural network.
- The Reverend and the Submarine
It starts with components of intellectual designs that go into self-driving car and how its software processes external objects. Then it goes to robotics revolution and SLAM (simultaneous localization and mapping) problem. To illustrate the problem author retells the story of missing submarine Scorpion and search for it, linking it to multistep Bayesian processing. Then author also links it to various applications from investing to medical diagnosis.
- Amazing Grace
This is about languages both natural and programming. Author suggested that we had 2 revolutions – one for each language: programming in 50s and natural language now, going from purely algorithmic construction to more probabilistic one. With this development computers had been achieving parity with humans in speech recognition. Author looks back at history to find tipping point of this development and for some reason uses story of Grace Hopper. Then he moves to history of computer software development from compilers to speech recognition.
- The Genius at the Royal Mint
It starts with the use of coin toss in sports and then moves to discussion of probabilities and anomalies, that allow identifying cheating. As illustration author uses Newton and his tenure at royal mint when he tried to fight money debasement and clipping by using Trial of Pyx – when accumulated over year samples where tested by the jury of experts. This method was not completely effective due to variability and Newton failed to fix the problem. After that author trying to demonstrate that contemporary statistical method would easily handle this challenge.
- The Lady with the Lamp
Here author moves to Florence Nightingale and her role in improvement of healthcare system, and use of statistical methods to achieve this. Author even calls her “The Mother of Evidence-Based Medicine”. It follows by illustration of contemporary problems and discussion of how AI could fix a lot of them.
- The Yankee Clipper
Somewhat unexpectedly author starts this chapter by cautioning against excessive enthusiasm for AI and pointing out at its dependence on assumptions that could easily lead to greatly diverse solutions produced with the same mathematical tools due to slight variance in assumptions. As illustration author uses data for Joe DiMaggio and some other examples, concluding at the end: Bias IN – Bias Out. As the final word author suggest that AI, even if it is unbiased and does better job than humans in many areas, should not be allowed to make decisions on its own, but should rather be used as quality of decision multiplier for humans.
MY TAKE ON IT:
It is generally correct description of ideas and math in foundation of AI, but I think that idea of AI working always under human control generally not realistic. AI is too fast and takes into account too many factors for humans to understand what, leave alone how it is doing something, so any detailed control is just not feasible. At the same time AI is just a tool, which is not conscious and therefore has no objectives or ability for self-direction. In short Self-Driving car would drive one anywhere, processing more information in seconds than this person could process in lifetime. However this car would never decide to go on car trip across the country for fun. In short I believe that AI will take over all routine and semi-routing jobs that human do for living both blue and white color. However since humans are the only self-directing entity in humans created environment, they will always decide “Where” and “When” to drive, even if reasons “Why” will always be sketchy, leaving to AI complete control over details of “How”.