Dual-Constraint Optimization of Mouse V1 Modeling: Integrating Sparsity and Excitation-Inhibition Balance for...

The primary visual cortex (V1) is central to mammalian visual processing and provides an important substrate for studying cortical computation and its relationship with artificial vision models. The Allen Institute's generalized leaky integrate-and-fire (GLIF)-based mouse V1 model (MV1M) is among the most biologically realistic large-scale cortical models, yet its ability to reproduce experimentally observed neuronal activity remains limited. Inferring high-dimensional synaptic connection weights in MV1M is particularly challenging due to expensive circuit simulations and the lack of tractable gradients. Furthermore, how biologically realistic vision circuit models align with artificial vision models remains insufficiently understood. To address these challenges, we propose structural sparsity and excitation-inhibition (E-I)-balance-inspired adaptive sequential neural posterior estimation (SE-IB-ASNPE), a scalable two-stage inference framework for MV1M synaptic weight optimization. The proposed framework first identifies informative sparse connectivity patterns to construct an effective prior space and then performs sequential posterior estimation guided by E-I-balance-inspired biological constraints. By integrating structural sparsity with biologically informed priors, the SE-IB-ASNPE reduces the effective parameter dimension and simulation cost while preserving circuit plausibility. We evaluate the optimized MV1M by comparing it with 32 ImageNet-1K pretrained deep neural networks (DNNs) across architectural families using layerwise representational similarity analysis (RSA). Results show that the optimized MV1M achieves stronger representational alignment with DNNs than both randomly initialized and original V1. Within several architecture families, MV1M-DNN peak RSA exhibits architecture-dependent directional tendencies relative to ImageNet-1K Top-1 accuracy, including positive tendencies in VGGs, ResNets, and ViTs and a weak negative tendency in EfficientNets (ENets). Although randomly initialized DNNs already exhibit nontrivial V1-like representations, ImageNet-1K pretraining on average increases their alignment with the optimized MV1M, suggesting that supervised visual learning contributes to brain-like representational similarity. The analysis with shape-dominant stimuli further reveals more structured representational geometry in optimized MV1M compared to randomly initialized V1, indicating improved shape-selective population representations beyond anatomical connectivity alone.