Video: How Much Custom Modeling Do AI Platforms Allow?

Learn more about video AI at Streaming Media's next event.

Watch the complete video of this presentation from Streaming Media West, T203. HOW-TO: Compare & Contrast Platforms Offering AI Video Insights in the Streaming Media Conference Video Portal.

Read the complete transcript of this clip:

Jun Heider: Training's a very important thing, and I see a lot of people are leveraging these services out of the box right now. The models are kind of closed.

Microsoft lets you train some things, such as face, and correcting the transcript, within your own account. Any of you ask at Google, good luck. AWS does let you upload your own faces, but if you were to want to add new labels, that's a closed door, and the way I think about it is, if I was malicious, and I wanted to say hot dog or not hot dog, and just have completely random images that weren't that, I could screw it up for everybody else, cause you're sharing all this training.

Lots of these vendors will allow you to create, upload training data, but it's kind of in a sandbox. Some of them do let you turn things on for anonymous training data. For instance, one of the services, I think, allows you to do anonymous training data for audio transcripts.

For the most part, there are not that many training options right now with some of these services, but I feel that going forward, that may change, because I had this example from somebody I was talking to at a previous conference, where he had a team create this big, robust taxonomy for all the content in their media library. He spent a lot of time doing it so that things could be found, categorized well, etc. Then, he had AI tag all this content, and now, he had all this metadata around all the content.

The problem was, his users that were trying to find the content were searching for things completely outside of the labels from AI and the taxonomy that his team had created. That's not cool.

In his case, he could have created a custom model based on search analytics: What are the terms that people are using to find that content within my media library, so that I can then create a custom model that has classifiers that match the things that people actually search with?

If you want to go down that path, Clarifai allows you to create custom models, and so do Valossa, Hive, and Adobe Sensei, when it actually comes out. Adobe says that they're going to allow you to do custom models.

Going forward, the more we use these services, the more we're probably going to be creating our own custom models that are specific to our content library.