Top-down feedback in deep neural networks leads to functional differences during audiovisual integration.

Incorporating top-down feedback into deep neural networks creates hierarchical architectures that produce human-like visual biases during audiovisual integration.

• Why it matters: Understanding the computational role of top-down feedback is crucial because it is a widespread feature in the brain that influences neural activity, yet remains underrepresented in artificial models.
• What they did: A hierarchical recurrent ANN model was developed to simulate top-down feedback, and its impact was tested on an audiovisual integration task across different network configurations.
• The result: Models mimicking human brain architecture exhibited a light visual bias similar to humans without impairing task performance, highlighting top-down feedback's role in shaping neural computation and behavior.