Intrinsic stabilization of synaptic plasticity improves learning and robustness in artificial neural networks.

Synaptic plasticity, defined as the ability of connections between neurons to strengthen or weaken over time, is essential for learning, yet unregulated plasticity can lead to instability. Inspired by biological feedback mechanisms, we propose a framework that stabilizes synaptic changes in artificial neural networks through top-down signals derived from the network's own output. The model, termed intrinsic Top-Down Stabilization (iTDS), augments traditional supervised learning by introducing a slower, top-down signal that tracks the network's output over time and modulates synaptic updates. Instead of directly influencing sensory inputs, these top-down signals guide synaptic plasticity by gradually aligning with ongoing network activity. We demonstrate that this simple mechanism improves training efficiency, enhances generalization, and increases resilience to noise perturbations. The approach is evaluated across recurrent, feedforward, and reservoir networks on 16 synthetic and real-world tasks, including temporal signal encoding, synthetic data categorization, handwritten character recognition, and image classification. Our results clarify how network activity influences the alignment of top-down signals with supervised learning, providing testable predictions for how feedback projections can stabilize learning in biological networks.