Video: Best Practices for Developing Machine Learning Algorithms for Video

Learn more about machine learning and AI at Streaming Media's next event.

Read the complete transcript of this clip:

Josh Gray: As we've gone after a lot of the projects where we applied techniques with large data sets, you know, you tend to see some of the kind of similar impediments. One is that data quality on the front end is often a challenge. It's a common misconception that you can just take a mess of data and throw it at an AI engine and magic will show out the other side. That's just not true.

We find that making sure that you're clear about what's in your datasets that you're going to be using to train, what are the semantics of the different attributes and the properties that you're going to feed into, the model development is really important. Being very crisp in what you want out of this thing. Do you want a classifier with keywords? Do you want rankings? Do you want comparative outputs? Understanding what are the signals of value that you're trying to train for and then how those relate to the attributes of your datasets, then making sure that throughout that stack that you're driving for clarity and definition, and don't just assume that it's mess in one side, magic out the other.

David Clevinger: I'm going to add to that and say that I think any strategy also has to include a data strategy, how you are training that algorithm, how you are helping it to understand what it is you want out of it.

I think my colleague over here is completely right. You need to know exactly what it is you're going to get out of it, but you need to know how you're going to get there as well. I think that also includes metadata flexibility. If you don't have a flexible metadata structure or someone on your team that understands how to build flexible, relational either hierarchical or relative metadata structures, it's going to be difficult for you to manage that on an ongoing basis. I think you have to have a training strategy going in.

Nadine Krefetz: Is there metadata-as-a-service?

Josh Gray: As we've gone after a lot of the projects where we applied techniques with large data sets, you know, you tend to see some of the kind of similar impediments. One is that data quality on the front end is often a challenge. It's a common misconception that you can just take a mess of data and throw it at an AI engine and magic will show out the other side. That's just not true.

David Clevinger: I'm going to add to that and say that I think any strategy also has to include a data strategy, how you are training that algorithm, how you are helping it to understand what it is you want out of it.

Nadine Krefetz: Is there metadata-as-a-service?

David Clevinger: Well, that's a great question. Companies like IBM might be building a metadata-as-a-service solution that you might be seeing in the near future, but there are other companies that do this already of course.

Then they do it in a variety of different verticals. It's typically done at the vertical level. You can find healthcare metadata companies. You can find financial services metadata companies. It doesn't really exist for M&E in a completely structured way because everyone is a little bit different. Sports versus movies versus what-have-you. But you could certainly use some existing products to build a metadata service.

Well, that's a great question. Companies like IBM might be building a metadata-as-a-service solution that you might be seeing in the near future, but there are other companies that do this already of course.