Training oscillatory FitzHugh–Nagumo neural networks with STDP: Dynamics and performance

Here we demonstrate for the first time that a neural network consisting of oscillatory FitzHugh–Nagumo (FHN) neurons can be effectively trained using spike-timing-dependent plasticity (STDP). Unlike phenomenological models such as the LIF model, where spike generation is governed by artificial threshold-and-reset conditions, the oscillatory FHN neurons in our network exhibit spiking activity that arises naturally from their own intrinsic dynamics. A crucial feature of the architecture is the time-delayed coupling between neurons, which provides a more biophysically realistic model of information transmission. Training the network on an elementary visual image classification task reveals that STDP drives the self-organization of a functionally significant and sparsified synaptic structure. Our analysis indicates that the learning dynamics are highly sensitive to the synaptic strengthening rate but robust to the weakening rate of STDP. An optimally tuned network achieves a level of accuracy that confirms the significant learning capability of this architecture before performance degrades due to overtraining. These results suggest that networks of oscillatory neurons with local plasticity and conduction delays are a promising foundation for developing interpretable, adaptive, and energy-efficient neuromorphic systems.