The Fast Track to AI with Javascript and Serverless

What is the work that you are doing today?

A lot of our work involves three threads. We're very active in serverless. And a lot of that is actually looking at legacy world challenged platforms and helping transform them into systems that can run in a serverless way. There's a lot of organizations that have realized that the economic operating model for cloud and particularly around serverless starts to make a lot more sense. So we're helping with a lot of that movement to serverless and also greenfield as well. Increasingly, a lot of that work involves the upgrading of those platforms to have additional machine learning capabilities. That's about augmenting the business process with things like natural language processing and understanding how to extract meaning from text, some image recognition and so on.

What are your goals for this talk?

The goals are similar to the book. The talk is based on our book, AI as a Service. And we wrote that to help people get on board with AI services. I think there's a lot of myth and uncertainty around A.I. and Machine Learning. You don't necessarily need to have a team of data scientists and a PhD in Machine Learning because of the commoditization that we're seeing around this technology, engineers can be very effective for the right use cases. And so we're trying to educate and inform developers about how they can get on board with this technology. I suppose the other part of it is that data science is one thing and building machine learning models is another. But that doesn't equate necessarily to a useful product or even a useful internal platform. There's an enormous amount of work around the edges that has to be done in order to operationalize these machine learning solutions. So there's a much wider context around that that we're trying to help educate people on as well.

How do you assess the quality of these models?

It's a great question. And of course, it's not just trusting the results. There are all the other subtle issues that come with using people's models, like bias in the results and so on, which is a huge issue at the moment. I suppose the answer is it depends, which I guess is not a very satisfactory answer, but one that is if you're using, say, a more simple point solution such as Textract from Amazon, which allows you to feed in an image and get back all the fields from the image. Say you wanted to extract a passport number, that can be done with an API call and that's much more boolean. It's right or it's wrong. Whereas when you get into things like classification and so on, it can be a little bit grayer. So I suppose the other part of the skill layer is being able to transfer learning with these cloud services. So I don't necessarily need to start from scratch. I can take a service like Comprehend and cross train so I can do my own customization on top of what's already there without needing to start from scratch. The other piece of the picture is the ability to operationalize my own models that I've built. Typically that process would be to use something like TensorFlow initially to do your research and then something like Sage Maker for additional training and then operationalize through those services.