Thinking Like Artificial Intelligence: What's in a Photo?
First in a series highlighting human thinking vs. AI.
Posted September 22, 2017
Artificial intelligence (AI) can do some pretty amazing things, but there are some other things that are very simple for humans and incredibly difficult for AI. To understand, it's helpful to think about how we humans make decisions and the process we go through. It highlights how some tasks may be difficult to explain to a computer, what we take for granted, and how we might automate some processes.
This week, we look at the task of identifying what's in a photo. Karen Zack has a hilarious collection on Twitter that challenges us to think about how we identify what's in a photo.
I love these images because, on the one hand, we humans can tell the difference. On the other hand, we have to look pretty closely: The color, texture, and even shape of the objects in each group of photos are similar.
If you were creating artificial intelligence to detect the difference between, say, a labradoodle and fried chicken, what should it look for? Are there cases where that wouldn't work? For example, you might program your AI to look for eyes, but black spots on the chicken may trick the algorithm, as may eyes hidden by fur. It's a hard problem.
The truth is that current AI actually can tell these types of photos apart pretty well. You can see one excellent example of such an algorithm on Twitter at @picdescbot. This is an AI program that describes what is shown in a picture. It usually does a remarkably good job, but it doesn't always get it right.
And this all matters. We humans are very good at image processing, but it is a complex, difficult, and expensive task for computers. (This article gives a good overview of some of the technical details of how it works.) However, the fact is that AI is nowhere near as skilled as people at recognizing what it sees. As this technology becomes more present in our lives, understanding these differences becomes important. For example, self-driving cars need to process the world around them (i.e. deciphering how a wall or truck is an obstacle).
So next time you're flipping through photos, think about how you figure out where one thing begins and another ends, what things are in the picture, and what might make them confusing. AI may go about this differently, but reflecting on how we process information ourselves is helpful when considering how AI might do it. Understanding your thinking more deeply will give you some insight into the difficulty AI might have in doing the same task without the benefit of all your human knowledge.
Facebook image: ruigsantos/Shutterstock