Visual representation of different levels of abstraction along the mouse visual hierarchy

The functional visual hierarchy has been widely studied in primates but was only recently revealed in mice. It is thus unclear whether different areas along the mouse visual hierarchy also represent distinct levels of visual abstraction. To address this question, we carefully designed a visual stimulus set consisting of natural images of diverse objects, synthetic textures, and corresponding noise from a range of families. We acquired and analyzed a novel electrophysiology dataset with ~13,000 neurons recorded across six visual cortical regions along the visual hierarchy in awake mice using Neuropixels probes, while presenting those images. By evaluating the performance of decoding tasks at three different levels (category, subcategory and exemplar levels), we evaluated visual information representation with different spatial granularity, i.e., whether a region prefers fine-grained features or coarse-grained global information. We found that the decoding accuracy of areas V1, LM and AL were consistently higher than areas RL, PM and AM for different levels of tasks, suggesting a potential separation of two processing streams. Within each stream, the early areas perform better in decoding exemplar-level image information, indicating that they are better at local feature extraction. On the other hand, higher visual regions perform better in categorization of texture families and objects, suggesting an abstraction of global information. Interestingly, decoding performance was significantly higher for textures than for natural objects, which implied a limitation of natural object representation in mice as compared to primates. To explain the change of representation granularity, we compared the representation of different visual areas with the pre-trained artificial neural networks (ANN) and observed an increased similarity between neuronal and ANN’s representations across convolutional layers. The overall results shed light on visual information representation at different levels of abstraction in the mouse visual hierarchy and provide a framework to study visual information transformation in mice.