From Machine Learning to Machine Understanding: Pieces, Interventions, and a Weak Self

Machine learning has become very good at learning. It is not yet good at understanding: models fit their training distribution and then fail on the interventions a person would handle without effort. We argue that understanding is a distinct stage that comes after learning, and that it should be built deliberately rather than expected to emerge from more data. We define understanding as linking invariant pieces to previously learned pieces, together with how each piece varies, so that answers follow from structure rather than from more data. We then describe how to get there, following the route children take from early heuristics to correct rules: learn candidate pieces, design interventions on which they disagree, let the world answer, keep only what stays invariant, compose what is kept, and abstain outside it. The main open problem is a weak self: a functional, consciousness-like self-monitor that notices when its pieces no longer cover the situation and, eventually, builds the missing piece. Position. The field should treat machine understanding as the stage after machine learning, and build it on purpose: a system understands a domain when it holds invariant pieces, knows how each one varies, has licensed them by interventions the world answered, and knows where they stop holding