causal
Rules Are Underdetermined by Data
When deep learning makes an error, it is not behaving randomly but learning a systematic, lawful rule that simply differs from how humans see the universe—because rules are underdetermined by data (like '2,4,6,8' could continue many ways), and deep learning's output-node-driven computation favors interpolation over the rule-seeking extrapolation that humans are built for.
causalpending
Speaker
Gary MarcusEvidence Quote
“the rules are under-determined by data... There are multiple rules. An easy example is if I say two, four, six, eight, what comes next? It could be 10, but it could be something else”
Created: 6/13/2026, 12:27:05 AM
Your notes
Loading notes...