It’s time for a Machine Language reality check

Gartner recently came out with its top strategic technology trends report for 2018. Highlights are here. Gartner tossed out a lot of red meat for people to chew on, but when the topic turned to artificial intelligence, one section stood out:

“Although using AI correctly will result in a big digital business payoff, the promise (and pitfalls) of general AI where systems magically perform any intellectual task that a human can do and dynamically learn much as humans do is speculative at best. Narrow AI, consisting of highly scoped machine-learning solutions that target a specific task (such as understanding language or driving a vehicle in a controlled environment) with algorithms chosen that are optimized for that task, is where the action is today.”

The reference to near-human capable AI has a lot to do with the steady stream of intelligent robots and computers that Hollywood has created since the 1950s, and science fiction writers and comic book artists going back to the 1920s. But this type of AI, known as general artificial intelligence, is decades away. A young robotics researcher at MIT told me earlier this year that she doubts she will see something like this in her lifetime. So, don’t count on R2D2 or a Cyberdyne Systems series T-800 showing up on the plant floor to take care of business anytime soon.

That brings us to Gartner’s claim that narrow AI is where the action is today. I partially agree.  The issues I have with this claim are:

  1. Narrow AI is not all machine learning (ML).
  2. ML is not ready for prime time, at least when it comes to industrial applications.

Machine Language reality check

Machine Language reality check Allow me to explain. While machine learning is a huge buzzword in all areas of tech, few companies can afford to make investments in the talent and applications and hardware required to do ML effectively. Further, all flavors of ML require enormous sets of data to be trained. We’re talking tens or even hundreds of thousands of images or inputs or whatever.

So, while a company like Google or Amazon can make the necessary investments to train an AI to tell a blueberry muffin from a chihuahua, it will be a lot harder for a midsized shipping company to train an autonomous forklift to navigate around an obstacle in a busy warehouse.

It’s worth noting that ML isn’t the only game in town. An older flavor of narrow AI that uses programmed routines or algorithms is still huge in industry. Examples include an industrial oven at a manufacturer that regulates temperatures based on inputs such as humidity and the grade of raw materials being used, or a mining cart that drives along a preprogrammed route and regulates speed according to weight and downstream readiness indicators. It’s not as sexy as ML, but it’s available now and can be used in all kinds of IoT-driven industrial applications.   

Endpoint: We’re approaching peak hype for machine learning. Three takeaways:

  1. The promise is huge, but the reality is ML is still too expensive for most industrial companies — the investment in time and money is significant, the tools aren’t ready, and the talent isn’t available unless you have really deep pockets.
  2. Be on the lookout for hand-waving vendors and experts promising untold riches from ML — and dominant tech players creating the necessary platforms for other vendors to build or connect ML applications.
  3. Don’t disregard other types of narrow AI based on programmed logic or optimization algorithms – capabilities may be limited to basic processes or routines, but the tools are available and are relatively straightforward to implement.
This is an excerpt from a previous edition of the Priority Payload Report (PPR) newsletter.