March 8, 2018
The Next Phase Of AI: Beyond The “Hype”
The Next Phase Of AI: Beyond The “Hype”
I was recently watching two luminaries in the area of artificial intelligence talk about where AI is at this point. It wasn’t really presented from the point of view that “one algorithm is better than another algorithm," but rather more in terms of how someone in business can make sense out of this new machine age we’re entering.
One of the more surprising points came from a speaker named Andrew Ng, a former chief scientist at Baidu who ran the company’s Artificial Intelligence Group and is now an adjunct professor at Stanford and a chairman of Coursera. Ng stated that 99% of the commercial uses of machine learning today come from what we call supervised learning.
Supervised learning represents one of three types of machine learning systems today. It’s when you have an event or object and a machine is not only told what that event or object is but then also learns to recognize it. You might say that autonomous driving or medical image recognition is a good example of supervised learning. Ng himself and a group of his students at Stanford built a system to recognize whether a head of lettuce was ripe or not – a project that all started with his students walking around with their cellphones, taking pictures of heads of lettuce and labeling them accordingly for the benefit of a machine’s understanding.
Next, we have reinforced learning, a type of machine learning that became most famous when systems were built that could learn, by themselves, how to play games - from Space Invaders to the game Go to Chess. In fact, the only thing these systems ever really need are the rules of the game. Once a system has such a foundation of rules, it can essentially take it from there.
Finally, there is what’s referred to as unsupervised learning. An example of this type of machine learning might be Google “learning” to work through millions of photos on its own in order to recognize and identify a particular image such as a cat.
The positive sentiment about supervised learning and its phenomenal progress over the last 5-7 years was echoed in this same presentation by Yann LeCun, a computer scientist who is the head AI scientist in Facebook’s research division. LeCun claims that the progress supervised learning has achieved, in substantial degree, comes from essentially two components: One is the ever-faster GPUs in auto specialized chips that are good at running neural nets and the billions upon billions of computations it takes to train those networks. The other is an ever-increasing availability of data to train these systems on.
The Promise Of Things To Come
Ng and LeCun may see that supervised learning is where a vast majority of commercial applications are today. Beyond today, however, there are many exciting evolutions of machine learning for tomorrow – these are not “experiments” (as some admittedly can be), but rather practical applications that can shake up a variety of industries.
Let’s take medical imaging recognition, for example. The current technological state of affairs is such that a system essentially has to be taught, “This image depicts cancer” and “This is not cancer.” Over time, the system is able to identify from a given photo whether a lesion is cancerous or not.
Yet, the real promise from here is that a person with a reasonably priced smartphone will be able to take a picture, upload it to the cloud, get it analyzed and receive the results of that analysis for a nominal cost.
As this technology progresses and becomes more widely distributed, the next true stage of machine learning could go a long way toward improving healthcare for more people in the world.
Are we there yet? No. It can take significant time for one system to analyze a variety of narrow problems to understand or rules to follow before it can solve a problem, play a game or perform analysis. My personal suspicion is that it might take billions of those systems, in turn, to build a system that we consistently recognize for its usefulness through the world.
Optimistically speaking, we will probably get to that point…even if we’re not quite there yet.
What you could say we are experiencing in machine learning is a new phase. We have been through the “hype” phase. We have also seen a system’s ability to learn and demonstrate broadly applicable tasks. From here, we will likely be entering a phase in which we have more sensors on a widespread basis – and in turn, as we are able to collect a large amount of data on a cost-effective basis, we should see exponential growth in a variety of useful systems. In addition, as a result of that exponential growth in computational capacity, we should also see systems built with the next layer of capabilities.
What Tech Skills Today Are Needed For Tomorrow’s Systems?
In practical terms, it may surprise you to learn that my expectation is that the technology skills that should be called upon to “help” these types of systems will actually be good old-fashioned software development skills in areas such as Java, Python and C++. I foresee a variety of candidates needing a solid familiarity with how to manage a large amount of data, including how to identify specific data in relation to a specific subject.
Now, some may believe that learning certain company systems will take an enormous effort. There will be substantial effort required, to be sure. However, for all their complexities, an understanding of the developmental core algorithms behind these systems may be less of a challenge than you realize.
Why? A variety of companies appear to be so intent on accelerating the learning curve for their systems that they are willing to give away all kinds of tools in order to induce people to learn their platform. Consequently, an individual who has just learned their system will essentially get “locked” into using the platform for the foreseeable future. We’re talking about companies like Amazon, Google, IBM and Microsoft.
As we pass from one phase of AI to the next, the idea that machines are “coming to take our jobs” is becoming ever more remote. If anything, machines will transform a number of roles and create new opportunities for those who are ready and eager to evolve with the times.
Of course, recognizing such new opportunities isn’t always obvious. That’s why it makes sense to partner with a technical recruiter that has over 30 years of experience like Roy Talman & Associates. Talking to us today about planning the next direction of your career path, especially as we’re on the cusp of seeing a big leap in the development of systems, may make for the best kind of timing.