by Babak Hodjat
Republished from Venturebeat, July 24, 2017
The resurgence of artificial intelligence in recent years has been fueled by both the advent of cheap, available mass processing capacity and breakthroughs in AI algorithms that allow them to scale and tackle more complex problems. Interestingly, this recent trend is reminiscent of the personal computing revolution of the ’80s, when cheaper and more available computing became a catalyst for mass “computerization” of numerous industries. Much like AI today, computers and computerization felt cutting edge and new, so companies were setting up computing departments and computerization task forces. By the standards of those days, we are all computer specialists today.
Adoption of computers didn’t come about overnight. Decades ago, there was high demand for computerization, but its implications for each industry were not clear. People sensed computers were important but weren’t 100 percent sure in what way. We had to go through a whole process of development and discovery, and, as a result of computer experts working hand in hand with domain experts over the course of 15 to 20 years, computers and specialized software were developed to suit different needs.
We’re following a similar path with AI. We’re now at the point where AI is often siloed in specialized departments and where C-suite players intuit how important AI will be but might not be sure how to approach it. Common questions today include “What is AI?” and “How can it help my business?”
How AI can help businesses
Let’s look at online content first, specifically website optimization. Most people now are familiar with conversion rate optimization (CRO), where site operators try to maximize conversions by testing new ideas for design, messaging, user experience, and more. AI can make this process more effective by orders of magnitude.
We need to figure out how we’re judging the AI’s solutions and define the world in which it operates. For this example, we judge success by increased conversions (and we can choose whether that means leads or sales or whatever) and define the world as a particular website and the changes the AI can make to it (fonts, designs, colors, etc.). We can give the AI information like changes to try (dozens of messages and design ideas), as well as the ability to determine browser type or logged-in status so the AI can also start segmenting users.
What happens with this approach can be staggering. The AI can find compelling combinations of designs and the audiences those designs resonate with. It can do this by leveraging genetic algorithms, effectively breeding fitter and fitter generations of designs that create children that convert more effectively and repeating the process as the AI bends toward more optimal configuration.
It’s important to note here an important aspect of this approach that fits a general definition of AI: It’s autonomous. The operator sets parameters and goals, but the AI decides the combination of ideas, always trying to find a better answer and better results against that goal.
There are many more such examples of successful AI-enablement in diverse industries ranging from finance and trading to health care and even agriculture. In all cases, some form of the steps noted above need to be taken, and these decisions cannot be made in isolation. This is a collaborative process that requires domain experts and AI practitioners working closely.
What AI is — and isn’t
But let’s get back to the beginning here: What is the essence of this AI? Unfortunately I do not have an easy answer, and I sincerely doubt that there is one. For one thing, the definition seems to have changed through time, and the expectation keeps exceeding the state of the art. Rather than coming up with a strict definition, I think it’s actually more valuable to look at some examples to show how muddy the parameters can really be.
Is keyword search on Google considered AI? You might think that the technology behind web search is pretty straightforward, but even all the way back in the late ’90s, search engines made use of the A* tree search algorithm, a technique that was taught in AI textbooks.
How about Siri? Well, surely a conversational system is an example of AI. Or is it? In the case of Siri, many attribute intelligence to its humor in answering questions like the meaning of life, or being able to tell a joke. The reality is that this aspect of Siri is based simply on a randomized look-up table. In other words, the aspects people find most lifelike are actually just engineers programming one-liners.
What about self-driving cars? Here too, most of what is being tested on the roads today, as well as the self-driving car that won the original DARPA challenge, was almost completely engineered, used sensors instead of AI, and did not have any learning capabilities.
In other words, it’s hard to tell whether the algorithm itself can be defined as AI or not, and I think that’s not truly all that important. The important part is if the AI improves upon one or more measures as defined by domain experts. It’s important if the AI models and learns the domain it operates in and is able to adapt to new circumstances and expectations. It needs to function autonomously and get better over time, no matter what we call it or how pure it is, from a definitional perspective.
The real thing we need to envision is a world where AI, like computers and the internet, is omnipresent. Because that world is coming. It’s a world where AIs design themselves through evolved neural networks (this is already underway and showing promise in achieving state of the art results on benchmark problems). It is, in short, a vastly different world than the one we live in today. The definition of AI will continue to change. It will continue to become more ambitious. It will grow. And just like computerization, AI enablement will only be fully achieved once all of us can be considered AI experts by today’s standards.
And that day is coming.