Are businesses dedicating enough resources to AI innovation?
Sort by:
That example makes people naturally uncomfortable but the process is valid. We were looking for remote volunteering opportunities for our annual philanthropy day and found a wildlife conservation site where you can pour through pictures to identify and tag animals. The more people that tag them, the more the system learns to do the same. They have this AI out there looking for endangered species so they need humans to do that manual identification to train it. So it’s a much more comfortable example, but the reality is the same. You need people that can identify what's there to teach it over and over again.
It's all part of training it—and if you don't have the right trainers, curators and overseers then you get what you get. If you're not paying attention to it, it'll evolve in an unpredictable way.
The challenge with AI and ML is not the modeling. There are enough researchers in the country that are trained in AI models. The challenge is the data and inferring the data—which we call the knowledge. Think about ML and AI today as nothing but expert systems: You're basically bringing in a domain expert and pairing them with someone who can build a model.
The best option is someone trained in both cyber and AI, but there are only a handful of us who understand both and can build one quickly. If not, you have to bring in a data person and a security expert before you start building up. The hard part is getting your hands on the data—the freshness and comprehensiveness of the data are two different things. We can get the initial data set, but trying to continually get the variety you need with the velocity that you want is what breaks the models. That's the real crux.
I think it depends upon the context, especially when it comes to data protection. As a difficult example, if I'm trying to prevent child pornography, how do I train ML for that without actually having people see it, gather it and use it? It has to be tightly controlled.
Having had investigators who worked for me and unfortunately had to investigate stuff like that, in many ways I'd prefer that they didn't have to view that material because of the impact on them personally. So a machine that would minimize some of that pouring through images and give them some highlights so they're not immersed in it would be a good thing for the security investigator. But you have to deal with some slippery slopes to create the machine learning capability to do it. Depending upon the context of what you're doing and how you're doing it, you have this proportionality of things that you have to think through.