A single complexity proxy does not consistently order well-generalized checkpoints. The same metric can correlate positively at one layer and negatively at another, and even within a layer well-generalized checkpoints can have high or low complexity. RankMe activation curves illustrate this, and Appendix J presents analogous curves for the other six metrics. · CiteArk
A single complexity proxy does not consistently order well-generalized checkpoints. The same metric can correlate positively at one layer and negatively at another, and even within a layer well-generalized checkpoints can have high or low complexity. RankMe activation curves illustrate this, and Appendix J presents analogous curves for the other six metrics.
来源:paper:Appendix I, pp.26–27; Appendix J, Figures 40–46, pp.28–31