It’s just past mid-year 2018. Back in December 2017 and January 2018, many pundits were making their annual predictions about artificial intelligence. In fact, my colleague Michael Porter wrote about AI predictions earlier this year. Rarely do I notice anybody checking in on those predictions to see whether they are reality or not. So I decided to check on the predictions made in the popular site CIO.com. Thor Olavsrud, the Senior Writer for CIO.com, made his predictions in a boldly titled article “5 artificial intelligence trends that will dominate in 2018“. Let’s break down these predictions to see if they have made their estimated impact so far this year…
Enterprises will operationalize AI
His first prediction was “Enterprises will operationalize AI”. To me, that says that organizations have gone through the experimentation phase and are now deploying AI into their businesses. To be sure, many companies have started to operationalize their AI work. But, as I see companies that we work with in the AI space, operationalizing AI is still the exception and not the rule. Most of these organizations are still thinking about AI – where it can be used, how to get started, etc. Many have begun experimenting with their own AI initiatives, but are still in the early stages of experiments. Maybe the pace will pick up in the second half of 2018, but I think a broader push to operationalize AI still hasn’t started.
AI reality will lag the hype once again
Mr. Olavsrud’s second prediction was “AI reality will lag the hype once again”. I think this is still true half way through 2018. As I mentioned above, the hype – that companies are operationalizing AI – still lags the reality that I see. I think AI has made some significant progress in many different areas and reality is starting to catch up with the hype. But, like Mr. Olavsrud, I don’t think we’ll see reality catch up with the hype in 2018.
Bias in training data sets will continue to trouble AI
The third prediction was “Bias in training data sets will continue to trouble AI”. This has been a traditional problem with AI and will continue to be a problem. The reason? Lack of data is one major reason. What, you say, but aren’t we swimming in data? In some areas are have an abundance of data – IoT alone is expected to generate 500 Zettabytes of data per year (1 Zettabyte =1.000,000,000,000,000,000,000 bytes).
But there are multiple problems within the AI / Data space. First, organizing and storing all that data is extremely complicated. Second, the kind of data organizations use to make decisions and where AI can be helpful is often missing or incomplete or small enough in size to make it hard to train an AI engine. For example, assume you are a bank trying to determine if a customer will buy more banking services. If you have 100,000 customers, you first have to determine the type of data needed to make those predictions and figure out how to collect and store it. Next you need to train the machine learning system to evaluate and make predictions. Say you use 10,000 accounts to train the AI tool. With such a limited dataset, training data can be easily biased.
So, yes, training data bias is still here in 2018 and will likely continue into the foreseeable future.
AI must solve the ‘black box’ problem with audit trails
A fourth prediction was “AI must solve the ‘black box’ problem with audit trails”. This concept is a particular problem for AI. The idea is that machine learning algorithms “learn” the data, but are not very transparent about how that learning works (thus the ‘black box’). When the AI tool starts making predictions, it is often hard to explain just how the machine got the answer. Imagine trying to sell a prediction tool to your CIO when you can’t explain how it works.
Audit trails may be able to help with this problem. Audit trails can at least show the inputs and maybe the scoring of data as the machine works on the prediction. In terms of the market, I think most AI initiatives are still trying to get AI right and will loop back to the black box problem in the future, so I don’t think we’ll see a solution to this in 2018.
Cloud adoption will accelerate to support AI innovation
The final prediction was “Cloud adoption will accelerate to support AI innovation”. I think this is absolutely happening in 2018. The major cloud vendors (Google, AWS, Microsoft, IBM, Oracle, etc.) are all investing in cloud-based AI capabilities. Even as software like TensorFlow make it easier to run complex algorithms on smaller desktop machines, the amount of data processed and the complexity of the processing systems still demand larger computing environments. Cloud computing is a clearly an advantage, especially in the AI experimentation phase, over investing in on-premise hardware.
Overall, I think these five predictions were pretty good half way through 2018. I don’t think we’ll see a lot of operationalizing AI this year, but maybe 2019?