预注册复现检验“做梦”Hopfield网络
一名Reddit作者报告了对Fachechi、Agliari和Barra“做梦”Hopfield网络实验的预注册复现，使用NumPy实现，包含1,000个神经元及每个条件10个随机种子。报告显示，t=10时经验容量倍数为3.86，区间为3.79至3.93；祖先重叠为0.15的相关模式容量为9.06。该复现增加了匹配的均匀收缩对照，并已提交ReScience C，代码、数据和预注册材料均已公开。
新增的均匀收缩对照可直接区分“做梦”机制的作用与普通耦合收缩，公开材料也使研究能在正式评审前接受审查。
I re-implemented the dreaming (Hebbian-unlearning) Hopfield kernel J(t) = J(1+t)(1+tJ)^-1 from Fachechi, Agliari & Barra (Neural Networks 2019) in NumPy, at N = 1 000, 10 seeds per condition, with a pre-registered design frozen by git tags before each run.
Main result: dreaming at t = 10 multiplies empirical capacity by 3.86 [3.79, 3.93]; 9.06 on correlated patterns (ancestor overlap 0.15). I added a matched uniform-shrink control absent from the original paper, which separates "dreaming works" from "shrinking the coupling works". One registered condition (overlap 0.30) is reported as inconclusive because of grid censoring, not hidden.
Submitted to ReScience C (reviewers currently being assigned). Everything is in github.com/hypneum-lab/rescience-nh01 at tag `submission-20261031`, with a Software Heritage snapshot.
Why post here: I'm a one-person lab working evenings. I'd welcome critique of the control design and of the capacity estimator before the formal review lands.
#Hopfield network#NumPy#ReScience C#Hypneum Lab#Neural Networks