Deep Convolutional Neural Networks Outperform Feature-Based But Not Categorical Models in Explaining Object Similarity Judgments

Abstract

Recent advances in Deep convolutional Neural Networks (DNNs) have enabled unprecedentedly accurate computational models of brain representations, and present an exciting opportunity to model diverse cognitive functions. State-of-the-art DNNs achieve human-level performance on object categorisation, but it is unclear how well they capture human behavior on complex cognitive tasks. Recent reports suggest that DNNs can explain significant variance in one such task, judging object similarity. Here, we extend these findings by replicating them for a rich set of object images, comparing performance across layers within two DNNs of different depths, and examining how the DNNs’ performance compares to that of non-computational “conceptual” models. Human observers performed similarity judgments for a set of 92 images of real-world objects. Representations of the same images were obtained in each of the layers of two DNNs of different depths (8-layer AlexNet and 16-layer VGG-16). To create conceptual models, other human observers generated visual-feature labels (e.g., “eye”) and category labels (e.g., “animal”) for the same image set. Feature labels were divided into parts, colors, textures and contours, while category labels were divided into subordinate, basic, and superordinate categories. We fitted models derived from the features, categories, and from each layer of each DNN to the similarity judgments, using representational similarity analysis to evaluate model performance. In both DNNs, similarity within the last layer explains most of the explainable variance in human similarity judgments. The last layer outperforms almost all feature-based models. Late and mid-level layers outperform some but not all feature-based models. Importantly, categorical models predict similarity judgments significantly better than any DNN layer. Our results provide further evidence for commonalities between DNNs and brain representations. Models derived from visual features other than object parts perform relatively poorly, perhaps because DNNs more comprehensively capture the colors, textures and contours which matter to human object perception. However, categorical models outperform DNNs, suggesting that further work may be needed to bring high-level semantic representations in DNNs closer to those extracted by humans. Modern DNNs explain similarity judgments remarkably well considering they were not trained on this task, and are promising models for many aspects of human cognition.

Evaluating (and Improving) the Correspondence Between Deep Neural Networks and Human Representations

“…More recently, it was shown that human shape sensitivity for natural images could be explained well for the first time using deep neural networks (Kubilius, Bracci, & Op de Beeck, 2016), which now constitute a near essential baseline for emerging models of human shape perception (Erdogan & Jacobs, 2017). A follow-up to our own previous work (Peterson, Abbott, & Griffiths, 2016) showed that important categorical information is missing from deep representations (Jozwik, Kriegeskorte, Storrs, & Mur, 2017).…”

Visual and semantic representations predict subsequent memory in perceptual and conceptual memory tests

“…Interactions among these three research domains have not been as intimate as one would hope. The domains of object vision and semantic cognition have been getting closer, both by the fact that semantic cognition researchers often examine the nature of natural object representations at both visual and semantic levels Martin et al 2018), and also that both domains are relying increasingly on the use of advanced neural network models to reveal statistical regularities in object representation (Jozwik et al 2017;Devereux et al 2018). However, the episodic memory domain has been somehow disconnected from the other two, partly because it has tended to focus on broad categorical distinctions (e.g., faces vs. scenes) rather than in the component visual or semantic features (Lee et al 2016).…”

Skeletal descriptions of shape provide unique perceptual information for object recognition

“…Instead, modern computational approaches to object recognition emphasize image statistics 39 or hierarchical feature extraction operations such as those implemented by convolutional neural networks (CNNs) 40,41 . Yet, without explicitly invoking any skeletal description, these models match human performance on object recognition tasks 41 , and they are predictive of both human behavioral and neural responses [42][43][44] . Even models that do emphasize global shape properties do so by describing the local properties of components parts (e.g., geons) and coarse, categorically-defined, spatial relations 12,45 , not a skeletal structure.…”