ML in Enterprise - soon to be disappointed?
TL:DR: Companies are setting high expectations for ML. Failures of ML projects will set a perception of ML as a failed technology.
Several companies are investing into AI and are working hard to advance it. Meanwhile, several others are using AI/ML as a filler for their non existent strategy. Marketing is using Machine Learning as that one weird trick. This can not last long.
Consider the following:
- Google and its angry customers
Google has angry customers in Europe, as YouTube has started serving the ads on hateful videos.
Google is a leader in machine learning for videos and employs some of the world’s best researchers. Despite this, the programmatic ad placement relying on ML fails in some cases. Not a huge surprise, except that press already making a big deal of this.
- Google Search results
The quality of Google search results have gone down lately. It has become rather easy for publishers of fake news to position their news at the top.
Google relied a lot on machine learning is starting to exhibits signs of failure in some cases.
- MD Anderson kicks out Watson after wasting millions of dollars on it.
This is a case of marketing getting ahead of reality and over promoting it. IBM has issued PR about Watson helping cancer diagnosis even before systems were on. A clear case of marketing gone bad.
There was also the failure of Microsoft Research bot named Tay that turned into being quite a racist. I made jokes about AWS Reknogition API, which seems to think a picture of hot dog as an inflatable thing.
Don’t get me started on security companies and their ML talk. Ask any infosec engineer and they will tell you all about this glorious vapor.
Granted many failures so far are in the consumer space and there are also some successes. For example, there is good progress on self driving front and other areas.
This is not to say that ML isn’t useful. ML is useful and companies need to continue focus on it. ML has huge long term potential. We need to continue the focus on training the models and do more testing. Its bit too early to position ML as a silver bullet for enterprises.
Perception is more important in any social system. Few more high profile failure will lead many stake holders in the enterprise to become suspicious of ML. We haven’t seen many failures yet in the enterprise because there haven’t been many projects in the enterprise that have gone to production. They will go into production in next 12–18 months and we will start seeing failures.
Enterprise vendors should call Machine Learning as an experimental tech for now and include with it plenty of warnings. I would also advice marketing to stay off of this. Focus instead on use cases your products can solve and treat ML as sausage making.
Thoughts?
