Pre-registered replication tests dreaming Hopfield networks
A Reddit author reports a pre-registered replication of the dreaming Hopfield network experiment from Fachechi, Agliari and Barra, implemented in NumPy with 1,000 neurons and 10 seeds per condition. At t=10, the reported empirical-capacity multiplier was 3.86, with a reported interval of 3.79 to 3.93, while correlated patterns with 0.15 ancestor overlap yielded a capacity of 9.06. The replication added a matched uniform-shrink control, and the study has been submitted to ReScience C while code, data and preregistrations are publicly available.
The added uniform-shrink control provides a direct way to distinguish the reported effect of the dreaming mechanism from ordinary coupling shrinkage, while the public artifacts allow scrutiny before formal review.
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