Deep Thinking by Garry Kasparov: The Education Of A Machine — Part 4/7
Garry Kasparov has written an extremely interesting and a very relevant book — discussing chess machines, Artificial Intelligence and the rise of our new computer overlords!
There are lots of observations about what separates a human cognition from the current crop of machines — as the book came out in 2017, Garry has seen all the developments incl AlphaGo.
I enjoyed the book — style and substance — at multiple levels. Bold, open, “fascinating, razor-sharp and provocative” as Andrew McAfee says in the back cover …
Note : It is instructive to listen to Garry’s TED talk , which provides a good backdrop for the book.
Side Note : Coincidently, as I was writing this blog, the documentary “Do you trust this computer ?” was released — adds more color to the book !! At the end of the blog, I will add some notes from that documentary.
Background:
There are many books and videos about the Kasparov-Deep Blue match including this one.
Deep Thinking by Garry Kasparov depicts Garry’s first person insights on the Deep Blue-Kasparov Chess games. That itself would have been sufficient to make this book very interesting. As Kasparov says “this is the first book that has all the facts and the only one that has my side of the story … it has been fulfilling to finally find the truth at the heart of Deep Blue…” I also have read Hsu’s “Behind Deep Blue,” so a view from the other side is illuminating.
In the process Garry also has tons of wisdom on AI as well as the Human — Intelligent Machine conundrum, which makes this book informative and relevant.
Kasparov’s Deep Thinking came out last year, full 20 years after the match, and it definitely has his wisdom not only on Deep Blue but also very good insights on the whole “AI is taking over the world” discussion.
From my pov, the book has five major themes, intertwined contextually in 11 chapters. As a short review, my goal is to tease out the themes that struck a code with me, as a guided tour to the book. You should definitely read the book, it is an excellent read.
1. Shall we trust the machines?
Garry is a little more cautious (in welcoming our new computer overlords) than Ken at the Watson Jeopardy match …
Or as Gavin says :
Kasparov ends the introduction with a very insightful observation — “ … we will face something new, something unsettling when we ride the first autonomous car or first time the computer boss issues an order at work. We must face these fears in order to get the most out of our technology and to get the most out of ourselves …”
Autonomous cars I can understand, …. a computer boss … almost all of us will draw a line way above that … but Garry had made his point …
If computers program themselves it should be OK, but we embed our biases and behaviors into the AI systems and that is what is more fearful …
2. The inevitability of Intelligent Machines
We are heading into an automated and artificially intelligent future at rapidly increasing speed and … “we don’t get to pick and choose when technological progress stops or where”
“Our attitude (towards AI) matters, and not because we can stop the march of technological progress even if we wanted to, but because our perspective on disruption affects how well prepared for it we will be. …
Each of us has a choice to make … to embrace the future and shape the terms of our relationship with new technology or resist and let others force the terms on us.”
…. Very well put Garry
Whether machines will take our job is a moot point. Jobs always have been automated — Kasparov talks about the elevator operator’s union, which was 17,000 strong in 1920 and they even paralyzed New York in 1945 with a strike !!
Another view point “Every profession will eventually be threatened by machines and it must, or else it will mean humanity has ceased to make progress” — This probably is Kasparov’s main thesis in this book
Garry progresses elegantly from Chapter 3 : Human vs. The Machine to his last Chapter, 11 : Human + Machine — a transition that is the main vein of the book’s thesis !
3. The idiosyncrasies of (seemingly!) Intelligent Machines
Kasparov points out lots of interesting shortcomings on the machines.
Who knows, they might come handy when (not if !) we face the apocalypse — so pay attention !!!
- 1968, Levy realized that computers are quite dangerous in tactical complications, clueless about the strategic plans and subtleties of endgame play. He would maneuver patiently, employing anti-computer strategy of “doing nothing, but doing it well”, until the machine would over extend and create weakness of it’s own position. Then Levy would clean up on the boards ! remember this when you encounter a computer !!! But soon, the computers started playing strong, consistent chess. But this strategy can be tries when we face the new computer overlords !
- Even Picasso got on the act “Computers are useless, they can only give you answers”!
- Ferrucci, the force behind IBM’s Watson, once mused “Computers do know how to ask questions, but they just don’t know which ones are important…”
- Kasparov mention’s Moravec’s Paradox a few times … A very relevant concept to keep in mind … Robots find the difficult things easy and the easy things difficult !
- When dealing with machines, we need to be aware of the psychological asymmetry and physical factors — the machine plays chess well, but also very differently. Keep that in mind if ever we have to encounter a war with the machines — The genetic footprint and DNA of the machines are very different from what we are used to dealing with fellow humans. The machine doesn’t care about styles or patterns or 100s of years of established theory — it is entirely free of prejudice and doctrine !
- Kasparov’s elaboration of how it feels to deal with machines is very illuminating … probably very rare insights … everyone should read them … be prepared … The Time article “The day I sensed a new kind of intelligence” where he says “ I could feel — I could smell — a new kind of intelligence across the table” is very eerie !
- Google has the mantra “ Data trumps everything” — This is not exactly how we build intelligent thinking machines. An interesting anecdote from the chess machines by Kathleen Spracklen as detailed in an interview. They fed grandmaster chess games to the chess program — everything seemed OK — the program developed pieces, launched an attack and promptly sacrificed the queen ! Usually grand masters sacrifice queen for a decisive and brilliant move ! So for the program “educated on a diet of GM games, giving up its queen was clearly the key” — it didn’t have the context to see when a queen sacrifice is relevant. Basically a mistaken correlation-casualty relationship
- A similar observation was made by Starman in the 1984 movie. He learns driving from his earth host and the first time when he sees a yellow light at an intersection, he speeds up ! Watching drivers, he learned the behavior - Green Go, Red Stop, Yellow go very fast.
- Obviously sacrificing a queen or driving through yellow lights need more context than the current machines have, if we let them learn only from data. Techniques like Reinforcement Learning can make this better, if we train them with enough context
- A very similar situation happened with IBM Watson during the Jeopardy match.
- The question was very straightforward. Ken Jennings raised hand first, saying “What is Only One Hand”. That was wrong. Watson had the correct answer “Leg” — Watson said “What is leg” Gymnast George Eyser was missing a leg. But Watson was wrong as well, the correct answer is “What is missing a leg”. Watson really didn’t know that having a leg is an oddity for a human — Jennings was right conceptually, but lacked data the human way; Watson had the right data, but not the broader context and human common sense — the machine way !
- Another one from the world of NLP. “The safe didn’t fit through the door but it was too narrow” — what does it refer to ? Now what if the sentence is “The safe didn’t fit through the door but it was too wide?” what does it mean now? The answer doesn’t lie in NLP or NER or any learned sentence structure but the common sense that “Wide things do not fit through narrow openings; door is an opening and safe is a thing” How would you teach a machine this ? via a knowledge graph ?
- Another one from the sentiment analysis (via Udacity) “I was lured to see this on the promise of a smart, witty slice of old fashioned fun & intrigue; I was conned” — again most machines would make mistake on this sentence
- Or what does “This fountain is not drinking water” mean?
- These are of course due to the ambiguity of the language- additional interesting examples from Udacity “A wise man is not the same as a wise guy. But a slim chance is the same as a fat chance”
4. Performance vs. Methods in AI
Performance is the ability for a machine to perform a task and method is the way it understands how to perform. The whole “Can it think?” question. As it turns out, we are still nowhere near achieving this …
Throughout the book, Kasparov has interesting discussions on performance over methods. And interestingly “when it comes to machine intelligence, we confuse performance — the ability of a machine to replace or surpass a human, with methods i.e. it has unlocked the secrets of human cognition and intelligence”
We seem too stuck in the “Don’t make it think, just make it work” attitude. In fact, as Kasparov argues “attempts to make machines that think like humans have failed, while machines that prioritize results over methods have succeeded”
Very true, even in Deep Learning, we embrace things that work, but we have very less understanding of …
Kasparov clearly articulates and dives into the two fallacies:
- The only way a machine will do X is if a machine catches up to human level general intelligence.
- If a machine can do X as well as a human, we have unlocked something profound about human intelligence.
As Garry Kasparov says in the introduction, “the tables have turned, as they always do in our eternal race with our own technology… [but] it turned out that making a great chess-playing machine is not the same as making a thinking machine on par with the human mind. Deep Blue was intelligent the way your programmable alarm clock was intelligent.”
In fact, nothing has changed even with AlphaGo; granted it is playing a game that is lot more complex than chess and is way more advanced than a programmable alarm clock, but still nowhere near a human mind. While AlphaGo is a toaster in steroids, the AlphaGoZero might be father out towards human mind — yet to be proven.
Another interesting observation “instead of a computer that thought and played like a human, with human creativity and intuition, they got one that played systematically… winning with brute number-crunching force.” Nothing has changed since then, AlphaGo included. Even AlphaGo really doesn’t know it is playing Go. The semantics is lost in the performance.
Of course the machines need not do the same way as nature to be effective — airplanes do not have flapping wings, helicopters do not have wings and wheels don’t exist in nature! When asked “Does Watson Think?”, David Ferrucci’s answer was “Does the submarine swim?”
I don’t think AlphaGo or even AlphaGoZero can teach the Go game or improve a Go player’s game. That needs a higher level understanding that computers still do not have. This is one of my pet peeves against the current AI domain.
As Kasparov laments “A chess machine that thinks like a human and loses to a Grand Master is not news; a machine that beats a world champion is new, nobody cares how it thinks” … So true, even in this age of Deep learning …
Kasparov hit the nail when he said “Applying context comes naturally to humans … “ probably because of the knowledge base of human common sense — maybe that is what is missing from the current AI systems !
I am working on a Semantic Go engine, which adds the common sense knowledge graph and the semantic knowledge of whatever it is doing … more later …
The theoretical & practical value of a semantic Go engine is much broader — teaching to social robots to …
BTW, the whole discussion of “Drosophila of AI” is interesting. It was chess in 1956 and Go till 2017 … probably semantic Go is next !!
The 2002 article in Dr.Dobb’s Journal is a good read. BTW, Am not sure how many remember Dr.Dobb’s journal !
Kasparov’s observation that “… how humans think is an excellent basis for relating how machines think and how they do not …” is very perceptive.
Humans have perception, feeling, remembering, but most importantly willing — wishes and desires. We also have a collection of common sense and a view of the world that has been developed over 1000s of years of evolution. Current machines lack that, but once we are able to embed that into a machine, probably it is game over or may be the game begins … !
A huge step in evolution would be when “machines move from surprising us with results to surprising us with the methods they use to find results.” Am not sure we are there, maybe Deep Reinforcement Learning.
Kasparov has a good understanding of AI and it’s evolution. He correctly observes “Deep Blue was the end; AlphaGo is the beginning …” very perceptive …. I agree !
5. Chess Players & Personalities
Throughout the book Kasparov touches upon the great players (including himself) and the surroundings. Great read!
We meet the inimitable Mikhail Botvinnik & Michael Tal “magician from Riga”, probably a cadre of chess players by themselves —” primus inter pares”, First among equals, as Kasparov calls them !
Note : I used to wait for the Bobby Fischer-Boris Spassky matches in the local papers, then play them and discuss the strategies ! Now I have moved on to Go matches by Lee Sedol ! Topic of my next blog …
His characterization of Bobby Fisher as “ideal challenger, but a disastrous champion” is interesting.
What separates Kasparov from other chess players ? “The willingness to take on new challenges, try new things and uncomfortable tasks” — probably an accurate self-observation and an excellent advice !!!
It seems once in the middle of a tough match, Mikhail Tal wandered into the Russian children’s poem by Korney Chukovsky “Oh, what a difficult job it was to drag out of the marsh the hippopotamus” …
And when he came out of the imagination the board became less complicated and executed a very intuitive the knight’s sacrifice; the move was praised by the press as being the result of deep thought, in fact the whole time Tal was thinking about the poem !
The undisciplined wanderings of mind, is a human trait — one machines probably won’t ever master !!
After the 8th game with Deep Blue, Kasparov talks about meeting actor Charles Bronson in the elevator - “They will never give you the chance” ! Oh, a bygone era !
6. And Finally …
Kasparov’s thoughts are very clear in the last chapter which might be the motif of the book —
It is healthy to be concerned about the directions of our technology …. But we might not have the foresight, imagination and determination to do what must be done. I might add “we might not get a chance or time or the visibility to act “ — as Garry points out future is always close enough to be ominous but never close enough to be in focus … extending when it is in focus, we might not see it at all, or we will see it but will not understand the significance !
The book ends very well:
Humans need machines to turn our grandest dreams into reality; And if we stop dreaming we might as well be machines !
Well said Garry !
