To Achieve Impact with Machine Learning, Talent Matters As Much As Technology

By definition a product means you can take it off the shelf, plug it in, and it works for everybody, without customization.

Definition (www.quora.com): Products are a collection of features and services packaged for sale as a unit without customization that solve a well-defined problem for a large market of similar buyers.

If someone is trying to sell you a machine learning software product that is supposed to work without customization (human oversight), my advice to you is to run in the opposite direction! Many of us have become dependent on the incredible software products delivered by companies like Google, Oracle and Microsoft. So it’s not surprising that businesses want to be able to rely on this familiar, plug-and-play solution for all of their AI and machine learning needs. But the very essence of what machine learning is, and why it’s so groundbreaking and valuable to business, is why it cannot be delivered by a product without human intervention.

How ML Works 

In order for ML to work, you need to get your data clean, configure the ML model, and then calibrate the model as it encounters data both in development and in production. This process of running data through the model and adjusting the model along the way trains the model to generate the appropriate results. One way to think about all of this is to consider the human brain. People are better able to think in general terms and draw conclusions when they have more data, more experience, and the accumulated wisdom to process that data. This is why parents have better judgement than their kids: because they have lived longer and have more data running through their brains.

The Red Bus

Imagine that you are walking your 5-year-old down the street and a red bus drives by. Your child says, “What is that?” and you say, “That is a red bus.” That child looks at the red bus and stores away in their brain the color, dimension, number of tires, number of windows, size, etc. Those are all data sets that the child will re-use the next time they see a similar vehicle. So then, when the next bus drives down the street, your kid may ask in a confirming fashion, “That’s a bus, right?” No matter what its color, a child still recognizes the bus.

ML technology can be built to recognize the red bus, but it won’t be able to generalize and recognize blue, yellow, black, etc. buses unless large amounts of data are run through the ML model and the model can be refined by humans. This ability to get the ML model to generalize is what separates the technology working in a vacuum from what can be achieved when the technology is built and monitored by true ML experts.

Silicon Valley tries to force everything into an off the shelf SaaS model, and Fortune 500 companies are enamored with using canned technology solutions, but ML isn’t there yet.  A stand-alone solution could predict the red bus and build simple neural networks. But when they try to apply it to all buses, it does not work, and it fails to capture the true promise of ML. Compared to the small universe of Fortune 500 companies working with ML luminaries who are maximizing this opportunity, corporations that have tried to rely on a software product for their machine learning approach have missed the bus, literally and figuratively.

The Pit of Product Despair

What makes the promise of ML so great is the capability to generalize and learn. Getting the technology to do that requires leveraging your data and some highly trained data scientists. There are roughly 20 recognized luminaries in the world – and perhaps a few of the most impressive doctoral candidates who studied under them – who possess the scientific expertise and technological sophistication to perform breakthrough work in the area of ML.

Yet, every CEO wants to assure investors and board members that they’ve got this problem solved. And most CEOs are hearing from their technology executives that they are effectively using ML already. In fact, the vast majority of organizations are unable to scale machine learning. So, ask your team these three questions:

1. How many ML algorithms do you have that are continuously working in your production environment?

2. How many of them are off-the-shelf products?

3. How many people do you have who are managing the algorithms and products that are in your production environment?

My guess is that your answer is, “Not many.”

CEOs must take the time to separate fact from fiction. There is no set-it and forget-it solution.  If you want to maximize the benefit of ML to your business, you are going to have to use a hybrid approach of technology that gets you 80% of the way there and people who will take it the rest of the way and maintain it.  Over time product software will offer better and better solutions, but like everything else in business, it won’t get the job done without people making it all work.

Read more: Machine Learning Is No Longer An Arduous Process