Fig. 1. The abstraction metrics and input representations.

a Two example classification tasks. (left) A classification learned between red and blue berries of one shape should generalize to other shapes. (right) A classification between red berries of two different shapes should generalize to blue berries of different shapes. b Examples of linear, abstract (left), and nonlinear, non-abstract (right) representations of the four example berries. c Schematic of the input model. d Schematic of the multi-tasking model. e Schematic of our two abstraction metrics, the classifier generalization metric (left) and the regression generalization metric (right).