New spiking neural network framework beats LIF with 10x sparser activity

A new paper from Alex Fulleda-Garcia and colleagues, published at the 2026 IEEE Neuro-Inspired Computational Elements Conference, introduces multi-timescale conductance spiking networks (MTCSN). Unlike traditional spiking neural networks (SNNs) that rely on simplistic phenomenological dynamics trained with surrogate gradients, MTCSN directly shapes the current-voltage (I-V) curve by tuning three distinct conductance timescales: fast, slow, and ultra-slow. This parametrization allows systematic control over neuronal excitability and can be implemented efficiently in analog circuits. The model naturally produces diverse firing regimes—tonic, phasic, and bursting—all within a single neuron type.

Critically, the authors derive a discrete-time formulation that enables direct backpropagation through time, eliminating the need for surrogate-gradient approximations. They benchmark feedforward MTCSN networks on the Mackey-Glass chaotic time-series regression task at the predictability limit, comparing against standard Leaky Integrate-and-Fire (LIF) and state-of-the-art Adaptive LIF (AdLIF) networks. Results show MTCSN outperforms both baselines in prediction accuracy while achieving substantially sparser activity—from both communication and computational perspectives. This suggests MTCSN is a promising building block for energy-aware temporal processing and neuromorphic hardware.