Logic Learning in Hopfield Networks

Synaptic weights for neurons in logic programming can be calculated either by using Hebbian learning or by Wan Abdullah’s method. In other words, Hebbian learning for governing events corresponding to some respective program clauses is equivalent with learning using Wan Abdullah’s method for the same respective program clauses. In this paper we will evaluate experimentally the equivalence between these two types of learning through computer simulations.

💡 Research Summary

The paper investigates whether two distinct methods for determining synaptic weights in Hopfield networks—traditional Hebbian learning and the analytical approach proposed by Wan Abdullah—are functionally equivalent when both are used to encode the same set of logical clauses. The authors begin by reviewing the Hopfield model, emphasizing its dynamical nature, the requirement for symmetric, zero‑diagonal connections, and the existence of a Lyapunov (energy) function that guarantees convergence to stable states. They then describe how propositional logic, specifically Horn clauses, can be mapped onto such a network: each propositional atom is represented by a bipolar neuron (state +1 for true, –1 for false), and the truth of a clause is enforced by constructing a cost function that penalizes unsatisfied clauses. Wan Abdullah’s method translates the logical program into a polynomial cost function, aligns it with the Hopfield energy expression, and directly derives the required connection strengths, including higher‑order (three‑neuron) terms. An illustrative example with three clauses (C ← A∧B, D ← B, ← C) yields a concrete weight table (Table 1).

Hebbian learning, on the other hand, updates synaptic strengths proportionally to the product of the activities of the participating neurons. The generalized Hebbian rule ΔJ_{i…m}=λ S_i S_j … S_m accommodates connections of any order. If the occurrence frequencies of events are governed by underlying logical rules, the Hebbian process should embed those rules into the weight matrix. Wan Abdullah previously showed that, under the condition λ = 1/n (where n is the order of the connection), Hebbian learning produces exactly the same weight values as his analytical construction.

Rather than comparing weight matrices directly—since interference and redundancy can cause numerical differences—the authors evaluate functional equivalence by running simulations. They generate random logic programs, convert each clause into Boolean algebra, assign a neuron to every ground atom, and initialize all connections to zero. Using Wan Abdullah’s procedure they compute the weight matrix, then let the Hopfield network evolve asynchronously until it reaches a stable state (minimum energy). In parallel, they simulate Hebbian learning on the same program by presenting the corresponding event patterns and updating the weights according to the generalized Hebbian rule. Both networks are subjected to 1,000 trials, each starting from 100 random initial states, with a tolerance of 0.001 for convergence. The authors record two metrics: (1) the ratio of runs that achieve a global minimum (i.e., a model of the logic program) and (2) the Hamming distance between the final stable states obtained by the two methods.

The experimental results, presented in Figures 1‑6, show that for networks of 40 neurons and for varying numbers of literals per clause (NC1, NC2, NC3), the global‑minimum ratio is consistently 1.0, indicating that every trial converges to a correct model. The Hamming distance is uniformly zero, confirming that the final configurations are identical regardless of the weight‑derivation method. The authors interpret these findings as evidence that Hebbian learning can indeed extract the underlying logical structure from event streams and produce the same solutions as the analytically hard‑wired approach. They also note that the networks never become trapped in sub‑optimal states, suggesting that the convergence time scales linearly with problem size.

The paper acknowledges several limitations. The logic programs used are relatively small and simple; scalability to larger, more complex clause sets remains untested. The choice of learning rate λ and the assumed event generation process are not empirically validated for realistic scenarios. Moreover, the impact of finite‑precision arithmetic on the derived weights, especially higher‑order terms, is not discussed. Despite these caveats, the study provides a solid experimental confirmation of the theoretical equivalence between Hebbian learning and Wan Abdullah’s method for encoding propositional logic in Hopfield networks, opening avenues for using biologically plausible learning rules to perform logical inference in neural hardware.