Preserving Linear Separability in Continual Learning by Backward Feature Projection | IEEE Conference Publication | IEEE Xplore