Abstract
The development of effective interventions for neurodegenerative disorders, such as posterior cortical atrophy (a visual Alzheimer's variant), remains to be a significant clinical challenge. We introduce a computational framework using convolutional neural networks (CNNs) as in silico models to simulate visual system degeneration and evaluate intervention strategies. By modeling controlled synaptic decay and comparing three distinct retraining approaches, random data (control), accuracy-based, and entropy-based, we assess impacts on classification performance and neural representation geometry. Our results demonstrate that accuracy-based retraining outperformed other strategies, maintaining model performance and preserving optimal manifold geometry during intermediate degeneration stages. This computational analysis supports prioritizing accuracy-targeted interventions for cognitive compensation. Our framework enables rapid evaluation of intervention efficacy while elucidating computational principles underlying neurodegeneration and recovery. This approach offers a platform for refining strategies to slow visual-cognitive decline in neurodegenerative diseases, bridging mechanistic insights with clinical translation.