跳到正文
Patterns· Greco·收录于· 6 天前

Robust inference via top-down feedback and neural variability in recurrent vision models

Robust inference via top-down feedback and neural variability in recurrent vision models

原文摘要

出版方登记 · OpenAlex

Biological systems leverage top-down feedback for visual processing, yet most artificial vision models succeed using feedforward architectures, calling into question the role of descending cortical pathways. Here, we trained convolutional recurrent neural networks on image classification in the presence or absence of top-down feedback. Models with top-down feedback exhibited remarkable speed-accuracy trade-off and robustness to noise and adversarial perturbations but only when trained with stochastic neural variability, simulated by randomly silencing units via dropout. The feedback-stochasticity combination reshaped representational geometry at the integration of bottom-up and top-down streams, reduced model activity dimensionality, and improved generalization to out-of-distribution inputs. Together, these findings uncover a dual mechanism for robust sensory coding, whereby neural variability prevents unit-level co-adaptation, while top-down feedback harnesses high-level information to stabilize network activity on compact low-dimensional manifolds. Our findings pave the way for biologically inspired vision architectures that move beyond feedforward designs by incorporating top-down pathways alongside stochastic regularization.

页面未能读取全文,以上为出版方登记的摘要;方法与图表见原文

来源:Patterns · doi.org