Research & Papers

Researchers propose Phase State Space Models for spiking neural nets

New 'Phase State Space Models' train spiking neural networks 40% faster without surrogate methods...

Deep Dive

Wilkie Olin-Ammentorp introduces Phase State Space Models (PSSM), a new approach that brings the parallel training power of state-space models to spiking neural networks—surrogate-free. The framework offers a novel interpretation of resonate-and-fire neurons, compatible with both real and spiking inputs, supporting parallel and recurrent execution, with clear links to hyperdimensional computing and biologically realistic features. The paper demonstrates a spike-compatible network that integrates STFT, recurrent memory, and attentional features in a single architecture.

Key Points
  • Phase State Space Models (PSSM) enable 40% faster parallel training of spiking neural networks without surrogate gradient methods
  • New approach unifies resonate-and-fire dynamics with state space architectures while maintaining biological plausibility
  • Implementation integrates STFT, recurrent memory, and attention in a single spike-compatible network

Why It Matters

Could dramatically accelerate development of energy-efficient neuromorphic chips for edge AI applications.

📬 Get the top 10 AI stories daily