Over the last couple of years I’ve become more and more interested in the ways that the research being done into Artificial Intelligence (AI) might be applied to production and post production. In this article I’ll be giving an overview of what AI is at this stage of development, and what technologies are being used. Later articles will cover immediate and future applications and implications.
One of the powerful way Artificial Intelligence ‘learns’ is by using neural networks. Neural Networks are trained with a large number of examples where the result is known. The Neural Network adjusts until it gives the same result as the human ‘teacher’.
However, there’s a trap. If that source material contains biases – such as modeling Police ‘stop and frisk’ – then whatever biases are in the learning material will be contained in the subsequent AI modeling. This is the subject of an article in Nature: There is a blind spot in AI research and also the praise of Cathy O’Neil’s book Weapons of Math Destruction that not only brings up that issue, but the problem of “proxies”.
Proxies, in this context, are data sources that are used in AI programs that are not the actual data, but rather something that approximates the data: like using zip code as a proxy for income or ethnicity.
Based on O’Neil’s book, I’d say the authors of the Nature article are too late. There are already institutionalized biases in very commonly used algorithms in finance, housing, policing and criminal policy.
Google have open sourced it’s Show and Tell model for automatically captioning images. This is an excellent example of how neural networks work: train the model with examples – in this case human captioned images – and then let it loose on new images. From the Venture Beat article:
Google trains Show and Tell by letting it take a look at images and captions that people wrote for those images. Sometimes, if the model thinks it sees something going on in a new image that’s exactly like a previous image it has seen, it falls back on the caption for the caption for that previous image. But at other times, Show and Tell is able to come up with original captions. “Moreover,” Shallue wrote, “it learns how to express that knowledge in natural-sounding English phrases despite receiving no additional language training other than reading the human captions.”
As the article points out, there are many more players looking to do the same thing. Imagine how much easier life would be in editorial if all the B-roll came in organized like this.