Superstitious perception by humans and convolutional neural networks

The advent of complex Hierarchical Convolutional Neural Networks (HCNNs) has led to great progress in the field of computer vision, with modern implementations of HCNNs rivalling human performance in object recognition tasks. The design of HCNNs was inspired by current understanding of how the neurons of the human visual system are organized to support object recognition. There are researchers who claim that the computations undertaken by HCNNs are approximating those of the human visual system, because of their high accuracy in predicting the neural activity of regions of the brain involved in object classification (Cichy, Khosla, Pantazis, Torralba, & Oliva, 2016; Khaligh-Razavi & Kriegeskorte, 2014; Yamins et al., 2014). However, there has been little investigation of how HCNNs and humans compare on other tasks that HCNNs have not been trained on. Our study compared the similarity of one HCNN, AlexNet, and humans on a superstitious perception task that involves falsely recognizing a learned object in the absence of strong evidence for its presence. We began by validating a new technique that quantifies human performance on the superstitious perception task. The first phase of the research revealed that human behaviour in the task is dependent on whether participants employed an active or passive task strategy. Next, the responses of our HCNN to the same images were analyzed in a similar manner. The results showed that HCNNs behaved similarly to humans in some ways and differently in others. Specifically, the classification images generated for the HCNN were similar to those derived from human participants, but the HCNN was also more consistent in its responses than humans. A second finding was that the differences in human participants classification images (created by adopting active versus passive strategies) could not be accounted for by simply altering the proportion of false alarm