Title: Is Meta-learning for Theorem-Proving One of the Keys to
Artificial General Intelligence?

Abstract: Within the class of cognitive architectures that leverage
uncertain logical inference as a key tool for creating and evaluating
knowledge and hypotheses, it appears a handful of hard technical
issues stand between the state of the art and AGI at the human level
or beyond.   One of these is neural-symbolic integration, i.e.
efficiently and usefully translating between the representations of
deep neural nets and other sub-symbolic pattern recognizers and
logical formalisms.    Another is meta-learning for inference, i.e.
inference guidance via recognizing patterns across previously done
inferences, and via inductive, abductive and deductive inference based
on these patterns.   I will describe some current work being done in
the latter direction, within the OpenCog cognitive architecture, using
a hypergraph pattern miner integrated with a probabilistic logic
engine to recognize simple patterns among commonsense inferences.   I
will also present some ideas about potential synergies between work on
inference meta-learning for commonsense reasoning with an AGI goal,
and work on meta-learning for reasoning in a more conventional
automated-mathematics context.