Skip to main content
. Author manuscript; available in PMC: 2020 Jan 6.
Published in final edited form as: Proc Mach Learn Res. 2017;70:3987–3995.

Figure 4.

Figure 4

Average classification accuracy over all learned tasks from the permuted MNIST benchmark as a function of number of tasks. Our approach (blue) and EWC (gray, extracted and replotted from Kirkpatrick et al. (2017)) maintain high accuracy as the number of tasks increase. SGD (green) and SGD with dropout of 0.5 on the hidden layers (red) perform far worse. The top panel is a zoom-in on the upper part of the graph with the initial training accuracy on a single task (dotted line) and the training accuracy of the same network when trained on all tasks simultaneously (black arrow).