AI In Action: Deep Learning Cracks Poker Code (Part I)

• (0:30) - What Is AI?

Welcome back to Mind Over Money. I'm Kevin Cook, your field guide and story teller for the fascinating arena of Behavioral Economics.

Since I am an investor in an exciting technology company you may have heard of called NVIDIA (NVDA), I often find myself in the position of having to explain to my followers and fellow investors "what exactly is AI" in a practical, right-now sense, and not some science fiction sense.

NVDA's type of computer chip, the GPU, is at the heart of modern AI R&D and they sell a lot of them not just for advanced gaming graphics but also to industry for applications in autonomous driving where Tesla (NASDAQ:TSLA), Toyota and Mercedes are customers.

NVDA also has a bigger business selling their processors to big cloud companies like Amazon (NASDAQ:AMZN), Google (GOOGL), Microsoft (NASDAQ:MSFT), IBM (IBM) and Alibaba (NYSE:BABA). In fact, NVDA AI chips power IBM's Watson.

If you want to learn more about NVDA's GPU chip technologies, see my video Get Your MPA in Deep Learning.

In the podcast that accompanies this article, I break down where NVDA will earn its $9 billion in revenues this year.

What’s the Difference Between Machine Learning and Deep Learning?

My objective today is to explain what the terms "machine learning" and "deep learning" mean and what their applications can be used for. Here are my layman's definitions, because I thought if I can't put computer scientist jargon into my own words, then I don't understand it well enough.

Machine learning (ML) is how computer scientists can train a system to recognize sets of data and their attributes well enough to compare/contrast and make decisions about what is selected or acted on when presented with new data.

Engineers speak of "training" computers to analyze, evaluate, and choose at various pre-determined decision points based on past recorded experience with a database.

Deep learning (DL) is an advanced subset of ML that takes everything a step further by creating new outcomes or "answers" that are not part of a pre-determined list of choices, or a database. The description of what is going on here is called "inference."

Disclaimer: I am not a computer scientist or engineer. I am simply an investor who is fascinated by technology that changes our lives, or that solves uniquely challenging problems, even as it invents new ones.

So I will not pretend to be an authority on AI, machine learning, or deep learning.

Consider this podcast an experiment... where I will attempt to highlight the unique aspects of these exciting technologies. It's such a big topic, I'll be doing this in two parts at least.

Because the science and the companies applying it could fill hours and I'd never get bored. I hope after this first episode that you feel the same way.

I’ll be reading from some articles on the work of the computer scientists who solved a particular puzzle and I will tell you where you can find the source articles and scientific papers, including the NVDA blog and podcast.

If I screw something up, or could have explained it better, I expect to hear from you: email me at KCook@Zacks.com

And if you are an AI expert who wants to be on the show to help clear the waters I might have muddied, let me know that too.

Hellmuth's Hell

Today's lesson begins with a poker blow-up...

I spent some time this past weekend watching YouTube videos of pro poker tournaments. Phil Hellmuth is infamous for not keeping his cool like Daniel Negreanu. Hellmuth gets emotionally upset and angry about two things:

1) opponents not playing rationally optimal (he shames them)

2) wild randomness (he blames bad luck as well as opponent's good luck)

See my recent video Gaming the Stock Market with Poker Smarts for more on the behavioral skill of Negreanu.

I'm not a very experienced poker player and I've never played high stakes. So I can't say I wouldn't get as emotional as Phil Hellmuth after losing $50,000 or $500,000. But I do have deep experience where this type of self-management does not work: trading and investing.

In the early 2000s I invented a probability training simulation called Probability, Risk, and Optimal Profit (PROP) to teach new traders how to over-ride their emotional patterns and cognitive biases.

The way I learned to trade, probability-based thinking and habits were the antidote to irrational, emotion-driven decisions. I used to teach my students this phrase...

"I'm just a surfer looking for the next wave, never blaming or hating the ocean for being itself."

After a weekend of thinking about Hellmuth and how I have spent the last 20 years trying to do the opposite of him in the stock market, I was thrilled to stumble across the NVIDIA AI Podcast on Monday morning in which host Michael Copeland interviewed Michael Bowling, the lead scientist behind the extremely rational and disciplined skill of an AI poker “player” named DeepStack.

You can learn more about it in my podcast, including how to get the 2015 Science paper from DeepStack's inventors at the University of Alberta.

Deep Blue Was "Brute Force" Computation

When you think of computers winning at complicated games with randomness, probably the first thing that comes to mind is IBM's Deep Blue tackling Garry Kasparov in a 6-game match in the 1990s.

That was a super-computer which could calculate all the possible decision points out to some end-game state and then choose the highest probability path each time. Billions of calculations per minute are what is known as "brute force" computing.

Other early applications of ML were email spam filters (no offense to the canned meat or anyone who enjoys it).

And ML computer vision is being used in factories and autonomous driving systems like that of Mobileye (OTC:MBBYF) which was purchased by Intel (INTC) this year for $15 billion.

Cyber-security companies like FireEye (FEYE) also use both machine and deep learning applications to detect and thwart network threats.

Solving Texas Hold'em

But compared to chess, poker is much more complicated puzzle to "solve" because it offers incomplete information: you don't know what cards the other player is holding.

Poker is not only a terrific laboratory to do AI R&D, it's also a great classroom to learn about the methods and achievements of AI scientists.

And that's because the heads-up limit version of Texas hold'em offers complexity of just 10 to the 14th power, whereas the no-limit version creates nearly infinite complexity of 10 to the 160th power!

For their first victory, the team at the University of Alberta used AMD (AMD) chips in their poker-player-destroying machine. For the new wrecking ball, they used NVIDIA's GeForce GTX 1080 series.

Check out my podcast to learn what the computer scientists from Edmonton mean by "solved" and how they taught their machine to use its "intuition."

Intuition may sound like a stretch to teach a computer but researchers have been perfecting a prerequisite to machine learning for decades. They call it CFR for counterfactual regret. Be sure to tune in to hear what that’s all about.

Disclosure: I own shares of NVDA, BABA, FEYE, and AMD for the Zacks TAZR Trader portfolio.

Kevin Cook is a Senior Stock Strategist for Zacks Investment Research where he runs the TAZR Trader service.