Hopfield Nets


patterns
$$\boldsymbol{\xi}^\mu \in \{-1,+1\}^d, \quad \mu = 1,\ldots,N$$
i.i.d. Rademacher; load ratio $\alpha := N/d$
weight matrix (Hebbian)
$$W = \frac{1}{d} \sum_{\mu=1}^{N} \boldsymbol{\xi}^\mu (\boldsymbol{\xi}^\mu)^\top$$
update rule (sequential async)
$$s_i \leftarrow \mathrm{sign}(h_i), \quad h_i = \sum_\mu m^\mu \xi^\mu_i - \alpha s_i$$
one step = one full random-order sweep of all $d$ neurons; overlaps $m^\mu$ updated after each flip
masked query
$$\mathbf{s}(0)_i = \begin{cases} \xi^1_i & \text{w.p. } 1-p \\ 0 & \text{w.p. } p \end{cases}$$
fraction $p$ of entries zeroed out

dimension $d$ 100
perturbation $p$ 0.20 fraction masked
max steps $T$ 30 update steps
trials $K$ 100 per point
converge after steps 10 unchanged

$\alpha$ range
[0.01, 0.30] 40 points
idle
recovery spurious no convergence

load ratio $\alpha$ 0.10 ($N = $ 10)
idle

jsi@berkeley.edu
james-simon
gScholar
@fakejamiesimon