We study the retrieval accuracy and capacity of modern Hopfield networks of with two-state (Ising) spins interacting via modified Hebbian $n$-spin interactions. In particular, we consider systems where the interactions deviate from the Hebb rule through additive or multiplicative noise or through clipping or deleting interactions. We find that the capacity scales as $N^{n-1}$ with the number of spins $N$ in all cases, but with a prefactor reduced compared to the Hebbian case. For $n=2$ our results agree with the previously known results for the conventional $n=2$ Hopfield network.

Close

PrevNext

View figure in article

• Memory
• Neuronal network activity
• Combinatorics
• Neural network simulations
• Random bond Ising model