Attention mechanisms in Spiking Neural Networks: From neural encoding to hardware implementation

Spiking Neural Networks (SNNs) offer a biologically grounded and energy-efficient alternative to conventional deep learning, yet have historically struggled to match the representational power of Artificial Neural Networks (ANNs) on complex perceptual tasks. Attention mechanisms present a promising route to closing this gap, but integrating them with spike-based computation raises fundamental challenges spanning representational incompatibility, temporal credit assignment, and energy preservation. This survey provides a systematic analysis of attention mechanisms in SNNs, tracing the full pipeline from neural encoding strategies through algorithmic design to neuromorphic hardware implementation. We categorize spiking attention into four types: rate-based, temporal, spatial–temporal, and spike-based, and analyze how each interacts with different encoding schemes and hardware constraints. Applications spanning computer vision, speech recognition, robotics, and multimodal sensing suggest that attention-enhanced SNNs can narrow the accuracy gap with ANN baselines while retaining the energy advantages of event-driven computation. We further examine the field through the lenses of theory, algorithms, hardware co-design, and community infrastructure, identifying the open challenges that must be resolved for attention-enhanced SNNs to mature into deployable neuromorphic systems, and articulating promising research directions toward that goal.