Pedro Domingos

We won’t find intelligence at the bottom of gradient descent.

You have a problem with gradient descent?

Yes, it’s an extremely inefficient way to learn.

Do you have an alternative approach to learning that is more efficient than gradient-based optimization?

The point is that we need to find one. AI research has become too narrowly focused on one paradigm. Gradient descent is a local optimum, and we need to get out of it. Do you really think the cortex learns by GD?

Q1: Does the cortex learn? Q2: if it does, does it do it by minimizing some objective? Q3: if it does, does it do it by estimating some sort of gradient information, or something to that effect? Q4: if it doesn't, then what type of (gradient-free) optimization does it use?