Thread by @daviddunning6

Thread: Spending time this week exploring overconfidence in AI systems doing machine learning. It’s a thing: AI systematically overestimates the chance its judgments are correct. Fascinating. Leading causes: 1/5

forbes.com/sites/kareemsa…

1.Machines have trouble when new stimuli range outside of what was in its training set. A system set up to tell cats from dogs categorized the researcher’s human face as a dog with 98% confidence! 2/5
jramkiss.github.io/2020/07/29/ove…

2.Machines are overconfident when trained on little data. Turns out a little learning is a dangerous thing, after all, even for your neural net. 3/5
api.deepai.org/publication-do…

3.Machines with high capacity “overfit” their predictions to the idiosyncratic quirks of their training data. When moving to new data, that misleads. Low capacity machines don’t have the wherewithal to do that as much. 4/5
arthurdouillard.com/post/miscalibr…

Solutions in general suggest mixing up and corrupting the training data. Throw in a few curves and make it less clean. Echoes what works in human education: Inject desirable difficulties into training. 5/5
bold.expert/desirable-diff…