Figure 4.

Deep convolutional neural network performance on the object recognition task. Performance (percentage correct) on the five-option object recognition task. Networks were fine-tuned on the five target categories; top-1 accuracy was computed. For the CORnets (left), performance of the feedforward architecture decreased with an increase in background complexity. For recurrent architectures, this decrease was less prominent. For CORnet-S, there was no difference between conditions. Error bars represent the bootstrap 95% confidence interval.