Why people fail at deploying Machine Learning models into production
First of all, we’d like to thank Madrid meetups, they are unwittingly responsible of this post. The launching of Machine Learning models into production has been the main discussion topic in the last meetups in Madrid. Unfortunately, it was not handled by the main speakers but it brought the interest while networking chats instead.
More and more companies face the problem of having hundreds of models that predict lots of things but… they are not useful neither even possible in a real “business” environment.
Reasons why ML projects never reach production environments
All resounding “ml to production” failures that we have seen in our work life share the same pattern: Wrong Born Definition or WBD in The Gurus’ jargon . Let’s dig into this concept through some great sentences:
- “I have a lot of data”: Do the first thing that comes to your mind with it
- “I have a lot of money”: Let’s do some Machine Learning, but please, the most complicated stuff that is in the market right now
- “I have both a lot of data and a lot of money”: We need to be a Data Driven Company, I give you the money, please don’t bother me again
Seems like not a nice scenario… This is the reality in 99% of Spanish companies. Has anyone asked himself:
- “What info do we need to increase our business growth?”
- “Is the data we have really useful for that purpose?”
- In case Yes:
- “Is the data available in a normal business operation?”
- “Do we have (or at least think about) the necessary infrastructure to handle it efficiently?”
- In case No:
- “Have you think about gathering other data?”
- In case Yes:
These are just a few examples of questions that every person in charge of making his company grow through Data should ask himself. Appart from ourselves, I haven’t seen people asking these questions in my life. If you know someone, please comment at the end of the post, we’d love to meet him/her.
Our experience bootstraping ML projects into production environments
We have successfully launched several Machine Learning projects into production and they are still alive and providing value to our customers (forecasting of electrical power in wind mill farms, user complaints…).
I think the reasons behind our success are more than one, but I want to remark the product lifecycle definition. If anyone wants to bootstrap a Machine Learning project into production it is compulsory to define a strong end-to-end roadmap (data availability, business goal, infrastructure…). After that there is yet a lot to do, but the most important work is done.
Just think about the questions we wrote in previous paragraphs. If you have them into consideration, you will be closer than 99% of Spanish companies in having a successful Machine Learning project, which provides value to your business and customers.
Let’s bring the Know-How to the Madrid Data Scientists
In the last meetup, a girl told us that she had to go to Barcelona to attend to the only “ml to production” course in Spain nowadays. Those interested in that training can access the course in the following link. However, my question is:
Why do we need to go to Barcelona (already being in Madrid) to this kind of training?
It seems clear that Madrid deserves a good training on these topic. We are preparing materials for that right now. If you are interested, please follow us on Twitter @TheGurusTeam or LinkedIn, we’ll keep you up to date of everything related to the training.