Fig. 2
Task performance and solutions. (a) The average test accuracy of 100 networks trained on a task where the predictive feature, size, changed by \(\alpha =0.5\). (b) The final test accuracies for various values of \(\alpha\), averaged over 100 networks per \(\alpha\). Error shades and bars correspond to \(95\%\) CI. (c, d) The model can solve the task using two different internal strategies. In one strategy, (c) a representative network learned to measure “largeness,” where it perceived the shapes getting larger (positive \(\Delta Z\)) and expected them to get larger (positive \(\theta\)). In an equally effective strategy, (d) another network learned to measure “smallness,” where it perceived the shapes becoming less small (negative \(\Delta Z\)) and expected “smallness” to decrease (negative \(\theta\)). Both solutions are equally valid for solving the task.
