Elastic Spiking Transformer Enables Runtime-Adaptive Gesture Recognition
A new architecture known as the Elastic Spiking Transformer has been created by researchers, designed for Spiking Neural Networks (SNNs) that can adapt its model width and attention heads during inference without the need for retraining. Drawing inspiration from Matryoshka-style representation learning, this model incorporates nested elasticity within the Feature Extractor, Spiking Self-Attention, and Feed-Forward blocks through granularity-aware weight sharing. This innovation enables a single universal model to balance accuracy and efficiency on neuromorphic platforms like Loihi and SpiNNaker, overcoming the limitations of existing SNN architectures that rely on fixed parameters and computational graphs. The method aims to facilitate energy-efficient processing of event-based sensor data for healthcare applications, including gesture recognition.
Key facts
- Elastic Spiking Transformer is a runtime-adaptive architecture for SNNs.
- It dynamically slices network width and attention heads at inference without retraining.
- Inspired by Matryoshka-style representation learning.
- Uses granularity-aware weight sharing for nested elasticity.
- Targets deployment on neuromorphic hardware like Loihi and SpiNNaker.
- Addresses rigidity of current SNN architectures with fixed parameter counts.
- Aims at energy-efficient processing of event-based sensor data for healthcare.
- Focuses on gesture understanding applications.
Entities
Institutions
- arXiv