Abstract: PURPOSE: Reading systematically engages the lateral occipitotemporal sulcus, at a site known as the visual word form area (VWFA). While the prevailing recycling hypothesis posits that the VWFA emerges through repurposing a pre-existing region for recognizing written words, the original function of this prototypic region and how experience shapes its representation remains largely unexplored. The object space model recently proposed that this area might have initially been responsible for representing word-related features in non-word objects that are proximate to words in object space, subsequently expanding its representational area through word training. In this study, we leveraged fMRI and deep learning neural networks to test this hypothesis, thereby shedding light on the origins and evolution of the VWFA.
METHODS & RESULTS: We initially tested the ability to discriminate between words in a widely used convolutional network, namely the pretrained AlexNet, which was trained with the ImageNet dataset. Surprisingly, we discovered that this network exhibited a significant ability to discriminate between words. Removing images containing words from the training dataset did not significantly affect word discrimination ability. Instead, we found that the pretrained network can effectively extract and utilize implicit, useful features from non-word images for word discrimination with deep dream methods. Furthermore, by training different networks with different sets of images, we found that the images that are closer to the words in the object space contain more word-related features. Taking our study a step further, we employed functional magnetic resonance imaging (fMRI) techniques to measure the brain response of seven subjects to 20 words and 80 non-word objects. The results revealed that objects close to words in the object space elicited stronger responses in the VWFA, bolstering the evidence favoring the object space model as an effective framework for explaining object representation in the human brain. Lastly, by systematically varying the correlation between word identity and task requirements, we found that task-irrelevant exposure hindered word representation and impeded word discrimination ability. Conversely, as the degree of association increased, both the representation area and word discrimination ability increased, suggesting that deep neural networks develop the object space in accordance with supervised rather than exposure-based learning rules.
CONCLUSIONS: This study provides compelling evidence that the VWFA area have initially functioned to represent word-related features in non-word objects that are situated close to words within the object space. The process of reading training further refines these area’s representations to fulfill real-life demands, aligning with the principles of supervised learning. Collectively, our research strongly underscores the object space model as a comprehensive and systematic framework for understanding the emergence and evolution of category-specific brain areas. This holds promising implications for future electrophysiological research, guiding exploration into the complex interplay between neural representations and learned categories.

Key words: Visual plasticity, visual word form area, deep neural networks, object space