notes-cog-ai-knowledge

Let's use machine learning to discover semantic knowledge. This is close to the original vision that led to machine learning but it seems to have been eclipsed by data analysis.

Black-box methods present a challenge but we can use them in a second-order kind of way: train a neural net and then, rather than directly regarding the neural net weights as deliverable, use it in some way within a larger procedure for discovering semantic concepts and rules. For example, one could (a) examine the neural net to find what it has 'learned' (explain it), (b) use the neural net to more quickly generate/test semantic concepts/rules (c) use the neural net to test whether candidate concepts/rules seem to be correct.