Machine Learning – Coarse to Fine

Because a friend of mine posted it on Facebook, I ended up watching this video, How Artificial Intelligence Learns From Biological Intelligence (by Evan Ehrenberg) that discusses some introductory machine learning topics. Along the way, the narrator mentioned that children start out with rather blurry vision, which improves with age; also, that 30% of our cortex is devoted to processing vision. So, I looked it up, and found out both the learning and amount data are true enough.

There’s an important lesson here: complex tasks are better learned through refinement. So, for example, if you wanted to learn something complicated and intricate like, say, predicting a time series. Then you’ll likely have more success training a neural net if you first feed it blurry and coarse data, rather than accurate and precise data. Then, you can tune the net by successively feeding more precise data.

How quickly should you improve the data? I’m not sure. We could extrapolate from the visual cortex, but I’m too lazy to do all the math. What to estimate? The amount of time and practice that an infant has spent seeing vs their visual acuity. [newborn is about 20/400, 6mo is about 20/25, newborn only sees grays, 1wk old sees red/orange/yellow/green, 5mo blue/violet], vs number of neurons that need traning [retina 150 million sensors, 140 million * 2 hemispheres ref]