MTSpark: Bio-inspired SNNs cut energy 2x while matching human-level game play
New spiking neural network approach achieves near-human multi-task RL with half the energy.
Reinforcement learning (RL) agents struggle to master multiple tasks simultaneously due to task interference and high energy consumption. A team from New York University Abu Dhabi (Putra, Devkota, Shafique) addresses this with MTSpark, a novel methodology that combines Deep Spiking Q-Networks (DSQN) with active dendrites—a bio-inspired feature that allows neurons to process inputs more selectively. MTSpark also incorporates a dueling network structure and task-specific context signals to dynamically form specialized sub-networks for each task, while exploiting sparse operations to drastically reduce computational energy.
Tested on three classic Atari games (Pong, Breakout, Enduro), MTSpark achieved scores approaching human-level performance: -5.4 vs -3 (human) in Pong, 0.6 vs 31 in Breakout, and 371.2 vs 368 in Enduro. Crucially, it did so with about 2x lower energy consumption than comparable state-of-the-art multi-task RL systems. This paves the way for energy-efficient generalist agents—robots, drones, or game AI—that can learn and switch between diverse tasks without retraining, making them practical for real-world deployment where compute and battery are limited.
- MTSpark uses spiking neural networks with active dendrites to create specialized sub-networks per task, reducing task interference.
- Achieves near-human scores on three Atari games while consuming ~2x less energy than state-of-the-art multi-task RL methods.
- Combines deep spiking Q-networks, dueling structure, and task-specific context signals for efficient simultaneous multi-task learning.
Why It Matters
Enables energy-efficient generalist AI agents that learn multiple tasks simultaneously, critical for real-world robots and drones with limited power.