Putting deep learning to work

Hype aside, deep learning can open up new sources of data for business analysis

After demonstrating discontinuous jumps in image recognition performance and defeating Korean grandmaster Lee Se-dol at Go, a game long resistant to computer mastery, deep learning has kicked up a swirling cloud of hype. And controversy.

On the one hand, serious folks are studying how to prevent a recursively self-improving super intelligence from seizing Earth’s reins from humanity. On the other, IBM’s “cognitive” marketing claims are rightly being called out as hyperbolic. I think much of the excitement derives from the tremendous strides deep learning has recently made in processing less-structured input, like images and voice, that relate to the way we perceive the world.

From a business perspective the challenge is to soberly assess the usefulness of new machine learning techniques for application in the near future. First and foremost, it’s worth remembering that you can tackle a very large swath of business problems with well-worn techniques in the areas of regression, classification and clustering. New methodologies, while enticing to data science practitioners, come with increased risk. It may be easy to build a model with a new technique, but tricky to operationalize it, or the technique could make strong assumptions about the nature of the data to which it’s applied.

That said, it’s worth considering how deep learning can drive business value. I believe the primary avenue is by adding structure to less-structured data.

Consider streams of video data captured from cameras in a clothing store. These data in their raw form are not that useful -- just colorful pixels in time and space -- but a human watching the video streams could make useful observations about the behavior of customers visiting the store. In essence, we can give structure to the raw video data by interpreting what we see.

Similarly, deep learning can turn raw sense-like data (which tend to have a strong spatial component that deep learning is particularly adept at representing) into a stream of structured data: the time, number, age and gender of customers entering the store, which customer cohorts (demographically) visit various retail displays and sections, which interact with staff, which make purchases, etc.

These structured data are much more relevant for data science modeling. You may want to do a form of predictive intervention, where floor staff are prompted to interact with customers that exhibit certain behavior or fall into demographic groups where staff interaction leads to a higher purchase rate. You may want to determine which aspects of retail displays lead to greater customer engagement.

This level of analysis, at the customer and employee behavior layer, is not appropriate on raw video data -- deep learning can serve as a bridge between these worlds.

Despite providing great leaps in image recognition, game playing, and voice recognition, deep learning is not a panacea. It still requires (like any machine learning technique) substantial human involvement in representing the problem domain and providing training data. So while it’s worthwhile for folks to consider the implications of A.I. advancement on the scale of the next few decades, those concerns are not particularly informative as to its application today.

Similarly, while calling the current state of machine learning or deep learning “cognitive” is a serious semantic stretch, the strides we’ve made do unlock new use cases and usable data, like those described above.

If you already have quality structured data relevant to the business problems and opportunities you’d like to address with machine learning, there’s no need to introduce deep learning. If you lack structured data but have access to less structured, sense-like data, consider deploying deep learning to bridge the gap.

Copyright © 2016 IDG Communications, Inc.

How to choose a low-code development platform