Abstract
Within the range of images that we might categorize as a "beach", for example, some will be more representative of that category than others. Here we first confirmed that humans could categorize "good" exemplars better than "bad" exemplars of six scene categories and then explored whether brain regions previously implicated in natural scene categorization showed a similar sensitivity to how well an image exemplifies a category. In a behavioral experiment participants were more accuracy and faster at categorizing good than bad exemplars of natural scenes. In an fMRI experiment participants passively viewed blocks of good or bad exemplars from the same six categories. A multi-voxel pattern classifier trained to discriminate among category blocks showed higher decoding accuracy for good than bad exemplars in the PPA, RSC and V1. This difference in decoding accuracy cannot be explained by differences in overall BOLD signal, as average BOLD activity was either equivalent or higher for bad than good scenes in these areas. These results provide further evidence that V1, RSC and the PPA not only contain information relevant for natural scene categorization, but their activity patterns mirror the fundamentally graded nature of human categories. Analysis of the image statistics of our good and bad exemplars shows that variability in low-level features and image structure is higher among bad than good exemplars. A simulation of our neuroimaging experiment suggests that such a difference in variance could account for the observed differences in decoding accuracy. These results are consistent with both low-level models of scene categorization and models that build categories around a prototype.
Introduction
Human categorization is rarely a binary process. While we can easily identify both a pristine tropical shore and a rocky, overcast coastline as "beaches," the former is intuitively a "better" example of the category. This concept of graded membership suggests that categories are organized around prototypes—idealized versions of a concept. Behavioral research has long shown that humans categorize these "good" exemplars faster and more accurately than "bad" or atypical ones.
This study investigates whether this behavioral advantage is reflected in the neural architecture of the visual system. Specifically, the researchers explored whether scene-selective regions like the Parahippocampal Place Area (PPA) and Retrosplenial Cortex (RSC) show clearer neural signatures for prototypical images, and whether this effect is driven by high-level semantic meaning or consistent low-level visual features.
Methodology
The researchers employed a multi-stage approach combining crowdsourced ratings, behavioral testing, and neuroimaging:
- Stimulus Selection: Using Amazon Mechanical Turk (AMT), the authors collected representativeness ratings for over 4,000 images across six categories: beaches, city streets, forests, highways, mountains, and offices. Participants rated how well each image exemplified its category on a scale from 1 (poor) to 5 (good).
- Behavioral Testing: Participants performed a rapid categorization task. "Good" and "bad" exemplars were flashed briefly to measure accuracy and reaction times under pressure.
- fMRI and MVPA: Nine participants viewed blocks of good and bad exemplars while in an fMRI scanner. The researchers used Multi-Voxel Pattern Analysis (MVPA) to train a classifier to decode the scene category from brain activity patterns in V1, the PPA, and the RSC.
- Computational Modeling: The team analyzed the "form" (via Gabor filters) and color statistics of the images to quantify visual variance and ran simulations to see if neural pattern clarity was a direct result of image consistency.
Key Insights and Results
Pattern Clarity vs. Signal Magnitude
The study's most significant finding is the dissociation between the clarity of a neural pattern and its overall strength. While "good" exemplars were significantly easier for the MVPA classifier to decode, they did not elicit a stronger BOLD (blood-oxygen-level-dependent) signal. In fact, in the PPA, "bad" exemplars evoked higher overall activity. This suggests that the brain may exert more effort—perhaps due to increased attentional demands or lack of repetition suppression—when processing atypical images, even though the resulting neural representation is less distinct.
The Role of Low-Level Features
Interestingly, the decoding advantage for good exemplars was present even in the primary visual cortex (V1). This indicates that what makes an image a "good" beach or forest is rooted in consistent low-level visual structures that early visual areas are tuned to detect. Image analysis confirmed that good exemplars are visually more similar to one another (lower variance in form and color) than bad exemplars.
Behavioral Performance
The neural findings mirrored behavioral results: humans were significantly more accurate (92% vs. 66%) and faster at identifying good exemplars. The simulation results suggested that the higher visual variance among bad exemplars is sufficient to explain the drop in both human performance and machine decoding accuracy.
Conclusion
These results suggest that the PPA, RSC, and V1 do not just process scene categories as discrete labels; their activity patterns reflect the graded nature of human perception. The clarity of our internal representations is heavily influenced by how well an object or scene conforms to a visual prototype. Future work may further investigate whether the increased BOLD signal for bad exemplars represents a specific "uncertainty" signal or simply the metabolic cost of processing more complex, non-prototypical visual information.
Summary updated