Recent advances in Deep convolutional Neural Networks (DNNs) have enabled researchers to create accurate computation models of brain activity. DNNs present an exciting opportunity to understand how the brain coordinates various cognitive functions, such as visual processing and object recognition—indeed, some DNNs can achieve human-level performance on object categorisation. In this study, researchers aimed to assess whether DNNs can match human behaviour on complex cognitive tasks. First, humans performed similarity judgements for 92 images of real-world objects, creating conceptual models using categorical labels to describe the images (e.g. “animal”, “eye”, etc.). Then, researchers used these labels, as well as image features (e.g. colour, texture, etc.) to train DNNs in object recognition. This enabled the researchers to compare human object recognition abilities to those of DNNs, given the same categorical and feature information. They found that categorical information, rather than object features, are much more important for producing similarity judgements between objects in both human participants and DNNs. These results provide further evidence for commonalities between DNNs and brain representations. However, human categorical labelling outperforms DNNs, suggesting that more work needs to be done if researchers want to create better computational models of cognitive processes.
Citation: Jozwik, K.M., Kriegeskorte, N., Storrs, K.R., & Mur, M. (2017). Deep Convolutional Neural Networks Outperform Feature-Based But Not Categorical Models in Explaining Object Similarity Judgments. Frontiers in Psychology.