Core Concepts
Layerwise complexity-matched learning enhances neural alignment in cortical area V2, leading to improved biological representation and generalization in object recognition tasks.
Abstract
The study introduces a novel layerwise complexity-matched learning approach to enhance neural alignment in cortical area V2. By matching task complexity with processing capacity at each stage, the model achieves better alignment with selectivity properties and neural activity. This methodology results in improved performance in predicting V1 and V2 neural responses, outperforming other architecture-matched models. Additionally, when used as a front-end for supervised training, the model shows significant improvements in out-of-distribution recognition tasks and alignment with human behavior.
Key points:
Introduction of layerwise complexity-matched learning for enhanced neural alignment.
Methodology focuses on matching task complexity with processing capacity at each stage.
Demonstrated improvement in predicting V1 and V2 neural responses compared to other models.
Application as a front-end for supervised training leads to better performance in recognition tasks and human behavior alignment.
Stats
The left panel shows that improvements in object recognition performance are strongly correlated (r = 0.57) with improvements in accounting for human recognition capabilities.
The right panel indicates a positive correlation between recognition performance and the ability to explain responses of IT neurons recorded in macaque monkeys.
LCL-V2 achieves state-of-the-art predictions of neural responses in cortical area V2.
Quotes
"We overcome limitations by developing a bottom-up self-supervised training methodology."
"Our layerwise complexity-matched learning formulation produces a two-stage model that is better aligned with selectivity properties."