The journey from supervised to unsupervised learning systems is going to require a patient approach, counsels AI strategist Chris Benson.
One of the promises of artificial intelligence is that one day it won’t be necessary for humans to inject themselves into the processing of data, either through the tedious task of labelling or else by providing a machine learning system with stated goals. Rather, the AI would be able to process huge amounts of data and teach itself how to arrive at desired outcomes. It would be able to apply what it learns in one area and transfer that knowledge to another area, bringing it much closer to thinking in a way that humans think.
What might the path look like to get to that point? We connected with AI Strategist Chris Benson – formerly Honeywell International’s Chief Scientist, Artificial Intelligence & Machine Learning within their Safety & Productivity Solutions strategic business group – to get his perspective on how AI development will play out over the next several years. Chris is a frequent keynote speaker at AI conferences and co-hosts the popular Practical AI podcast.
“With so much research going on and so much money being poured into this because of the potential upside, we don’t even know what new techniques may come out in the days and years ahead. But I will say that you will see a continuous stream of these advancements instead of one giant leap. I think it will be lots of little pushes in various directions to get us there.”
– Chris Benson, AI strategist
ANSWERS: What do you see as the most significant element of artificial intelligence development within the next few years? Conversely, what is the most overhyped?
CHRIS BENSON: I would say that the most significant element would be the shift to new training approaches. Current supervised learning approaches require an extremely labor intensive annotation of the training data in order to classify and categorize it, which inhibits scaling AI out through business operations.
I would say the most overhyped element of AI is that it will be pervasive and successful in everything we do, starting immediately. In my experience, many organizations out there that are attempting an AI strategy are failing to first implement an appropriate and comprehensive data strategy to support it. Data strategy is really a prerequisite to AI strategy, and you can’t really achieve success in the latter without first achieving success in the former.
ANSWERS: Can you talk to us about what an unsupervised learning system is and what advances need to take place to help those unsupervised learning systems proliferate and take some of the manual process out of things?
BENSON: Let me do that by contrasting it with what a supervised learning system is, because that’s really where most of the work in AI and the development of models is happening today.
To create a supervised learning system, you have to create a training set by labeling all the data. To label is to create annotations that classify your historical data and categorize that data so that what you are essentially doing is saying that your dataset represents the reality that you’re trying to train your model against. It does that by looking at a history of inputs that produced certain (and desired) outputs, and it trains against that set of data in a supervised context. That requires an enormous amount of manual labor that can take weeks and months to get prepped ahead of time.
A giant leap forward in artificial intelligence would be to move to the other side where you have unsupervised learning, and by unsupervised it basically means you’re going to take unstructured data and apply algorithms to it that doesn’t require that you to first create that dataset for training. There are all labeled and classified and categorized, and there is a lot of research going on in that space right now.
Traditionally, data scientists have had different types of techniques around clustering of data where an algorithm would take a lot of unstructured data and basically show that some of those bits of data point to results over here and some of them point to results over there. That is very useful, but it still has limited functionality in what you can apply that to.
One of the areas in neural network development where it’s being used is in what are called autoencoders. It will try to apply algorithms that reduce dimensionality of your data. What that really means is you might have many inputs going into a neural network model, and autoencoding will actually reduce that number of inputs into something a little bit more manageable. That’s a very simplistic way of describing it. That’s one area.
There’s a lot of research right now in terms of trying to find that because if we can get to widespread use of unsupervised learning systems or maybe new alternatives that we haven’t thought about yet, then that should enable us to tremendously scale out AI capabilities across organizations and accelerate the use of these technologies by removing the manual labor that the current supervised approach requires.
ANSWERS: What do you forecast is the timetable for when we might move to a greater advent of unsupervised learning systems? How close are we?
BENSON: There’s a lot of research in that area right now. I don’t really have a timetable in mind that I would have any confidence in at this point. It’s not just a world where you’re only looking at supervised and unsupervised reinforcement learning, which is a technique where you see your output and make adjustments to your model to make sure that it starts getting to a better and better place. There are lots of different techniques.
With so much research going on and so much money being poured into this because of the potential upside, we don’t even know what new techniques may come out in the days and years ahead. But I will say that you will see a continuous stream of these advancements instead of one giant leap. I think it will be lots of little pushes in various directions to get us there.
ANSWERS: How do you see artificial intelligence and similar technologies impacting the human capital needs and business strategies of companies in the next decade?
BENSON: Today, you tend to see very specific tasks in a narrow context that we are potentially replacing. For instance, a human might have done something (maybe on an assembly line) with an AI-enabled robot or something like that. I think over time what’s going to change is that, instead of just that swap out where you’re pursuing a given business strategy and you just say, “Hey, I can use AI on this task rather than a human,” you’ll start to see that businesses are accounting for these capabilities through their organizations and then their very strategy itself will change to accommodate that.
What you’ll see is that businesses will start analyzing their own operations: “Where do I want a human? Where do I want an AI-enabled technology? How can I use them together collaboratively to get farther than either could alone?” You might think of AI as supercharging a human in that kind of context, but we will have a working relationship with the AI. It will do certain things better than we will do and we will do certain things better than it can do, and we will work together to achieve the organization’s objectives.
ANSWERS: What innovations do you foresee in machine learning and AI germinating in the near future? How will they augment the human component?
BENSON: It’s really hard to predict. The thing that is driving this is that we have three exponential growth curves that are collaboratively driving innovation, both within AI and that AI can produce subsequently.
Those exponential growth curves are
- Computing capacity
- Availability of data
- Algorithm development
As all three are pushing exponentially it means that the advances we have in AI are really being made faster than any previous technology in human history. The very nature of what it means for humans to live, work and play is changing faster than our ability to keep up socially, culturally and economically.
You’re going to see across industries that products and services and bots and robots that are all AI-enabled are going to become pervasive. From this point forward we’re entering into that AI-human partnership that I mentioned before.
In medicine, there are examples that we’re already realizing today that are advancing. In medicine, a surgeon who is going through a very difficult surgery may be there directing a number of robots that are AI-enabled. They may have perception modules that are assisting the human in terms of being able to see and analyze what is there before them. It may be that rather than that surgeon needing to have a steady hand, they have a robot that is making a cut. The robot can do it at the thousandth of a millimeter in terms of precision. That’s just one use case in medicine out of a thousand.
For years now, machine vision has been used in agriculture. You have a large farm machine rolling across a field that might have millions of plants in it; you can analyze every plant’s needs and give it the right treatments that it needs.
This goes across just about every industry on the planet. I get asked variations of this all the time, and I’ve yet to come up with an industry that is not seeing change now and in the very near future because of this. It’s fundamentally changing not only what it means to work and in the innovations that we see there, but really in what it means to be a person, what it means to human.
ANSWERS: There’s a lot of interest in the technology but at the same time there might be a tendency toward wait and see. How would you coach an organization to remove any reluctance on investing in artificial intelligence?
BENSON: I would start with asking what the organization is trying to achieve. In all likelihood, its objectives are tied in some form to what their customers need, and what is the customer experience that the organization is trying to fulfill? I would look at the AI technologies that are available today and some of the more well-understood approaches (machine vision being a great one; natural language processing being another) that are becoming fairly common and ask, “Can one of these technologies better enable the customer experience that your organization is trying to create?”
If in a very practical sense you say, “This is a place where this new technology, this new AI toolbox has some better tools in it than what we were using before,” then that’s a great place to invest. It’s very practical, it’s a by-the-numbers kind of approach rather than looking at the technology as magic that you would like to harness.
That’s what I encourage people to do – treat your AI decisions just as you would any other business decisions driving your organization. If this tool is likely to work better than the previous tools that you have, then it’s a great time to invest. That is the organizational perspective.
There is also the individual perspective, which can be both from people in the organization and people who are those organization’s customers. From an individual perspective, the willingness to try AI or utilize AI in a productive way comes down to trust. You have to give people a reason to trust the AI they are either implementing or that they’re about to use. They need to know that it’s being implemented within an ethical framework just like any other prior tool would have been. I think it’s really key that there are research centers like the AI Now Institute at New York University that are trying to understand the social implications given the capabilities of this technology going forward.
In our new series, AI Experts, we interview thought leaders from a variety of disciplines — including technology executives, academics, robotics experts and policymakers — on what we might expect as the days race forward towards our AI tomorrow.