Convolutional neural networks (CNNs) not only recognize objects with high accuracy, but also acquire from images abstract statistical concepts such as numerosity and correlations. However, it remains unclear whether the CNN architectures implement inductive biases that mimic human biases in statistical judgments. In this paper, we examined whether VGG-19 models, a popular CNN architecture, that are trained to make correlation judgments from scatterplots display human-like biases. In comparisons between model predictions and human data, we found that there was a high correspondence between human biases and machine biases in VGG-19 models. Using explainable AI visualization with saliency maps to unpack the regions on which VGG-19 rely to make correlation judgments, we found that the late layers of the model tend to focus on regions similar to human participants’ fixation distributions as captured by eye tracking. We further demonstrate that such models were nearly sufficient to predict human data at an accuracy level rivaling the state-of-the-art model trained on human data in three large-scale correlation discrimination datasets. Our results suggest that VGG-19 models may employ strategies that are similar to those used by human participants for statistical judgments from visual graphs and, therefore, pave the way to address human cognitive biases in visualization-based statistical judgments through the lens of deep neural networks.