How to organize artificial intelligence efforts at work?

Question

Some context: Recently all kind of salesmen have been knocking on our company's door to provide their "artificial intelligence" expertise and projects suggestions. Some don't know the difference between words estimation and validation (really), some have extraordinary powerpoints and paint themselves as gurus of the field. Our management has gone with the hype and definitely we're starting some kind of project on "artificial intelligence" (meaning rpa with some machine learning possibly).

What is the best way to start when we don't yet know to what problem we want to apply all this and I'm worried it will lead to long expensive projects with meager results? What are the things to watch out for? Any good practical books or war stories out there?

Answer 1

I know what you mean. It can be difficult to parse between the hype and the application of AI. Although AI (specifically deep learning) can do a lot of things, this doesn't negate previous methods that work just as well or better in certain domains. Some times managers will hear 'AI' and think of giant networks when they actually just needed a simple linear regression for their problem.

That being said, here are some suggestions to help decide if AI is right for your project and some things to consider if so :

• Get some advice from a researcher. There are many AI consulting firms popping up in large cities that give companies hours of consulting time. If there isn't one in your city, go to AI meetups and meet the community, find educated people and ask them about your general questions. For more in depth advice, perhaps a local university can lend a hand.

• Understand your problem. Taking the time to know exactly what problems you are trying to solve is invaluable. Not only will it help new employees but if you do end up using AI and get consultation hours from an AI expert, it will save time and money explaining what you want to do.

• Know your data. First this is a check to see if you have the right data for AI and if you have enough of it. For example many problems are approached using supervised learning which requires having a lot of labeled data (Eg, think of 1000s of pictures labeled Cat or Dog depending on the image content). If you don't have or can't easily collect a large amount of labeled data, perhaps you can use an existing data set with similar data to help get you started. If that's not an option, then you probably won't like the alternative where you hand label the data or hire Amazon MechanicalTurks.

• Be prepared to fail at first. AI is not easy and is even harder when you can't just Google your questions because you are doing something no one has done before. It takes some time to understand your problem, your data, and what kind of models you should try.

• Do you have the infrastructure? If everything is going well and you have a working model on your local GPU, if your model is part of your product, look at how you might deploy it. Do you have your own GPU servers? Can you afford the cloud GPS that some companies offer? Do you have to learn a model for each user?