New Memristive ONNs Enable Autonomous Learning with Signed Weights

A team of researchers (Riley Acker, Aman Desai, Garrett Kenyon, Frank Barrows) has introduced a new neuromorphic primitive for oscillatory neural networks (ONNs) that overcomes a long-standing hardware limitation: the inability to implement signed (negative) weights in practice. Their design uses memristive edges with inhibitory couplings, validated through circuit simulations for an auto-associative denoising task. The key insight is that signed effective weights are necessary for anti-phase attractors to persist autonomously after training, enabling richer attractor structures beyond purely synchronous couplings.

This work directly addresses a gap between numerical Hopfield/Ising models (which assume signed weights) and physical ONN implementations, which often lack negative weights due to device and circuit constraints. By providing a practical route to inhibitory couplings, the approach supports phase-coded memories where anti-phase constraints are autonomously maintained. The results position ONNs as a viable architecture for continuous learning and inference, with implications for low-power neuromorphic computing and complex optimization tasks.