Curve Fitting

Judea Pearl famously dismissed the achievements of deep learning as "just curve fitting," and the phrase has been read as a sneer. It is not. For Pearl it is a classification. To fit a curve is to find a function that passes through your data points and generalizes to nearby ones—a magnificent thing to be able to do, which modern systems perform across spaces of staggering dimensionality. But a curve, however many dimensions it inhabits, is a creature of the first rung of the Ladder of Causation. It describes how observed variables move together; it says nothing about what moves what. You can fit a curve to the relationship between barometer readings and storms with exquisite accuracy and predict storms from barometers all day long, and you still cannot use it to decide whether smashing the barometer will prevent the storm—because that information is about mechanism, and the curve is about pattern. This is why Pearl holds that scaling makes a system better at the first