How Do Machines Learn? Part III: They Recover
Boxer George Foreman fought his last match in November 1997. Foreman had an astounding 30-year professional career where he went toe-to-toe with most of the greats of his era.
I remember watching his last fight on Pay-Per-View, just a few months after the disappointing Tyson vs Holyfield match with the infamous ear-biting incident. Coming off that disappointment, I didn’t know exactly what to expect in the matchup and, the truth is, today I can’t even recall whether the fight was good or not. I know that Foreman lost a controversial decision, and was interviewed at the end where he confirmed that he was retiring. He was on camera, talking about the so-called Lean Mean Fat-Busting Machine George Foreman Grill and the announcers asked him what he planned to do next. I’ll never forget his answer:
“Man, I’m just here trying to sell my grill.”
The line has stuck with me all these years, leaving me wondering if that is the destiny of every fighter who turns in his gloves and has (relatively) good health. I think of fighters as people on a mission, looking to prove themselves in a competitive and challenging sport that has routinely demonstrated that both body and mind matter. But old fighters have done their job and at some point have to move on. The same could be said about machines (both those that bust fat and those that don’t). They are built for a purpose, and usually we run a machine until it breaks or becomes obsolete.
But artificial intelligence is a new frontier for machines. In Part I of this series, we showed that the first component of machine learning is training. That is, to find some data that provides known information about some process or task that a machine can mimic and to develop mathematical algorithms to “learn” how to perform that process or task. In Part II, we saw that fighting is a key component of machine learning, as features and mathematical formulas compete for the top spots in the convoluted black box that is a well-tuned artificial intelligence algorithm.
So what happens next? What happens to the algorithm after the fighting is done and it has proven its worth and is the champion? Well, this is where the machines diverge from humans and the real work begins as machines recover.
So what does recovery actually mean here? Certainly, there are no fishing trips or beach lounging for AI. There is no need to heal because machines don’t need sleep. The recovery here is driven by what AI can recover for humans. Whether it is time, money, or convenience, the long term goal of AI is to help recover something for humans. So once we decide that the machines have adequately learned how to do something, we put it into deployment as we try to scale up and scale out the process so that it can recover something for humans.
Algorithms in this so-called “production mode” may still see some additional tuning as they learn how to deal with heretofore unseen circumstances. But the overall idea is that the machines should take over some new task. We are seeing this all over the place, with stories toting blockbuster results, such as the billions of dollars health systems are saving by deploying AI algorithms to identify fraud, predict patient relapse/readmission, or partner AI and wearable technology to create early warning systems for patients in need.
There are a growing number of instances where recovery leads to machines replacing humans, from manufacturing to food services to well-paying, white collar jobs, such as those in the legal area. But this doesn’t necessarily mean that the work is a zero-sum game, in that every bit of recovery machines make is at the expense of some other human loss. Indeed, many long-term plans for AI involve integration of humans and machines. Consider self-driving tractor trailers. It is unlikely in the future that a stable deployment of this technology means that there are no more truck drivers. Driverless trucks would be open to all sorts of threats, from both hackers and traditional thieves. Trucking is just one of many instances where the future appears to be headed to a blend of human intelligence and artificial intelligence working in concert. (There is a great recent piece in Wired all about this).
The most interesting thing here is that much of this is so new. We haven’t hit the end of any AI lifecycles and there really isn’t much known about what happens long term after deployment. The issue is in deciding when to replace one algorithm with another. We’ve certainly found this to be true in the financial services industry. This piece by Reuters shows that 95 percent of all ATM transactions are run using COBOL computer code that turned 60 years old this year. COBOL seems to be the George Foreman of computer languages, fighting off newer algorithms and continuing to provide humans with convenience and value. Imagine where we will be in 60 more years if a couple of the health AIs being deployed today turn out to be the next COBOL.
And that does it for the How Do Machines Learn? series. Thank you for following along and don’t forget to leave feedback on Twitter. You can find me @DrJasonBrinkley.
Commentary on the greatest boxer and/or computer algorithm of all time pretty much guarantee a reply.
Jason S. Brinkley, PhD, MS, MA is a Senior Researcher and Biostatistician at Abt Associates Inc. where he works on a wide variety of data for health services, policy, and disparities research. He maintains a research affiliation with the North Carolina Agromedicine Institute and serves on the executive committee for the Southeast SAS Users Group. Follow him on Twitter. [Full Bio]
Previous posts in this series:
- How Do Machines Learn? Part II: They Fight
- How Do Machines Learn? Part I: They Train
- Taste Testing Generic Drugs
- Halloween by the Numbers
- What Kills Kids?
- The Golden Age of Health Research Funding
- Does Living on a Prayer Work?
- The Opioid Data Crisis
- Income Lost from Snow Days*
- What the #$@&*! Is Blockchain?
- Opportunistic Research Opportunities
- Text Mining UFO Data: Little Green Aliens or Santa’s Elves?
- Should You Know Your Doctor’s Home Address?
- The Population Bullet
- The Unknown Unknowns of Missing Data
- Communicating Science–More Than Just Good Words?
- Counting Alabamas
- The Third World in Your Own Backyard
- The Unrealistic Gold Standard
- Does MACRA Signal the Beginning of the End for Medicare Claims Data?
- Think You Aren’t Extraordinary? Odds Are You’re Wrong
- Mapping by Words
- Are We Asking Too Much From Surveys?
- Making Better Comparisons
- What Kills Us?