Fig. 4: Accuracy comparison.
From: On the visual analytic intelligence of neural networks
a, b Average test accuracy for different model sizes (N value reported next to each point) for separate (a) and joint (b) training setup. Legend labels report the best average accuracy and the corresponding N value for each series. See Supplementary Note 3 for detailed values. c SNN (N = 128) separate training accuracy with min/max error bars plotted along with the human-level accuracy, for riddles ranked by difficulty level for humans reported in ref. 17, for which a difficulty index ranging from 1 (the easiest) to 45 (the most difficult) is assigned. d Test accuracy evolution over the first 20 training epochs with marked epochs when reaching human-level accuracy obtained in a different setup, described in text.
