多项式分析显示未训练网络在V1脑建模上并不优于训练网络
Analysis choices decide whether untrained networks beat trained ones at V1: a multiverse analysis of model-brain RSA
做脑建模和 RSA 的朋友值得注意：这篇把以前未训练网络更接近V1的结论拆解了，发现换一种归一化或距离度量结论就翻转。训练网络在更干净的角度分量指标上其实更强，还提醒你亮度对比度基线已经很难打败。
一篇 arXiv 论文（arXiv:2611.15533）重新检验了此前关于未训练卷积网络在 THINGS fMRI 数据上建模 V1 优于反向传播训练网络的结论。在预先设定的 ridge readout 下，未训练网络在每个池化网格上得分都低于训练网络（g = 4 时中位 r 差为 -0.043，5 个种子中 0 个为正），MSE skill 和逐折 r 也一致偏低。
作者进行了探索性的多项式分析，遍历 36 种单元归一化、距离度量和体素归一化的组合（28 种不同分析），发现 V1 对比的符号依赖于分析选择：5 种分析置信区间在 0 以上，9 种在 0 以下。原始的 correlation-distance RSA 与模型和体素模式对单位均值轮廓的共变性高达 Spearman 0.97（跨自身网络）和 0.79（跨 6 种 ImageNet 架构）；而基于中心化模式的未加权角度分量时，训练网络在 V1（自身网络及 5/6 ImageNet 模型）和 LOC（全部 7 个模型）上均领先。
此外，一个仅基于亮度和对比度的特征库在 ridge 下得分超过所有网络，加入网络特征并未改善预测。作者建议今后报告对刺激中心化的敏感性、分别报告各分量，并纳入小型多项式分析和简单基线。
Analysis choices decide whether untrained networks beat trained ones at V1: a multiverse analysis of model-brain RSA
Untrained convolutional networks have been reported to rival trained networks as models of visual cortex, and we had reported an untrained-network advantage over backpropagation at V1 in THINGS fMRI, together with a dependence on evaluation resolution. Under a pre-specified ridge readout, the untrained network scored below the trained one at every pooling grid (median r difference -0.043 at g = 4, 0 of 5 seeds positive), and also in MSE skill and per-fold r. In exploratory analyses, across 36 combinations of unit normalisation, distance measure and voxel normalisation (28 distinct analyses), the sign of the V1 contrast depended on the analysis (intervals above 0 in 5 and below 0 in 9 distinct analyses), and in our networks the resolution dependence of the original convention appeared in only one cell. Raw correlation-distance representational similarity analysis (RSA) covaried with the alignment of model and voxel patterns with the unit-mean profile (Spearman 0.97 across our networks, 0.79 across six ImageNet architectures); the advantage was absent from the component orthogonal to the mean profile, and the parallel component alone did not resolve it. On the unweighted angular component of centred patterns, trained networks were ahead at V1 (our network and 5 of 6 ImageNet models) and at LOC (all seven); the norm-weighted angular term was not resolved. A luminance-and-contrast bank scored above every network under ridge (point estimates), and adding network features under a shared penalty did not improve prediction. Intervals are stimulus-bootstrap intervals for three subjects, and 6 pre-stated predictions were not met. We recommend reporting sensitivity to centring over the evaluated stimuli on both sides, reporting components separately, and including a small multiverse and simple baselines.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include RSA, V1, fMRI, 卷积网络.
Source attribution: curated from arXiv cs.LG (@Nils Leutenegger).
