Can Information Flows Suggest Targets for Interventions in Neural Circuits Appendices

Figure 8: Visualizations of accuracy and bias flows for the smaller ANN trained on the synthetic dataset. Note how the most dominant accuracy flows arise from X3, which is the only bias-free feature in the dataset. In contrast, the largest bias flows arise from X1 and X2, both of which are heavily biased features. It is intuitively clear from these pictures which edges have the largest bias-to-accuracy flow ratios, and hence which edges would be the first to be pruned.