Spiking Neural Networks Cut AI Energy Use in Half While Mastering Multiple Tasks

Researchers have developed a novel approach to training artificial agents on multiple tasks at once while consuming far less energy than conventional methods. The technique, dubbed MTSpark, leverages spiking neural networks (SNNs)—a computational model inspired by how biological brains process information—to achieve energy-efficient multi-task reinforcement learning.

The Challenge of Multi-Task Learning#

Teaching a single agent to master multiple tasks simultaneously remains a stubborn problem in reinforcement learning. When agents try to learn several skills at once using traditional neural networks, they often experience task interference, where training on one task degrades performance on another. Meanwhile, the computational burden of running standard deep neural networks consumes substantial energy, making deployment of autonomous systems in power-constrained environments impractical.

How MTSpark Works#

The MTSpark framework enhances a Deep Spiking Q-Network (DSQN) by adding active dendrites—inspired by the branching structures of biological neurons that allow them to perform local computation. These dendrites dynamically activate to form specialized sub-networks tailored to individual tasks while the overall network maintains shared knowledge. The system also incorporates task-specific context signals that guide which parts of the network to recruit for each task, enabling sparse computation that dramatically reduces energy demands.

Performance and Efficiency Gains#

Testing on three Atari games revealed competitive results: the approach achieved scores of −5.4 in Pong, 0.6 in Breakout, and 371.2 in Enduro—approaching human-level performance across all three. Crucially, the system matched the memory footprint of existing state-of-the-art methods while consuming approximately half the energy, opening pathways for deployment on energy-constrained platforms like mobile robots and drones.

Implications for AI Agents#

The findings suggest that combining biological principles with reinforcement learning could unlock a new generation of "generalist agents" capable of adapting flexibly to diverse real-world environments. By dramatically cutting energy consumption without sacrificing task performance, MTSpark points toward more sustainable and deployable AI systems.