Artificial Intelligence & Plato: Teaching DNN "Forms"
I had a chance to hear Dr. Anh Nguyen present the findings of his team's research on deep neural networks and image recognition at PSU this week.
He's put together an excellent video on how neural networks can be easily fooled - worth the watch
The difficulty of teaching AI to correctly recognize an image reminded me of Plato's allegory of the cave from an epistemological standpoint – how Plato believes we come to know things through their "form." To Plato, there had to be a pure form of "treeness," for example, for us to look at hundreds of different plants and understand they are all trees, regardless of looking different.
Plato's understanding of our learning is not without flaw, but it is interesting to think that the confusion of neural nets is due their lack of innate forms. A neural net must be fed thousands of images of trees to begin recognizing one in a photo, but will it ever establish "treeness" for itself?
The talk also made me reflect on my three-year-old daughter. I didn't keep a record of how many times I pointed to a real tree, or photo of a tree, and told her "this is a tree," perhaps 100-200 times? She can now identify what is a tree and what is not with ease. How many times does a person need to be told what a thing is before they can identify it?