ad1f8bb9b51f023cdc80cf94bb615aa9-AuthorFeedback.pdf

Figure 1: (left) Interpolating between the binary and one-shot algorithm with γ. (center) Transfer enables faster learning on SplitCIFAR. Figures viewed best with zoom. MNIST with a similar network: In GNu, Online EWC achieves 33.88% & SI achieves 29.31% on 10 permutations of MNIST [3] while SupSup achieves 94.91% accuracy after 250 permutations (see Table 5 in [3] vs. Table 7 in our work). The method provided for NNs will not work with data that is common between tasks. PSP on SplitCIFAR achieves worse performance than BatchE (GG) with similar bytes.