Split-Gate Memtransistors for Energy-Efficient Adaptive Reinforcement Learning

Area of Science:

• Materials Science
• Artificial Intelligence
• Edge Computing

Background:

• Real-time, on-chip learning is crucial for AI edge systems in dynamic environments.
• Analog in-memory computing offers energy and latency benefits but lacks reconfigurability for reinforcement learning.
• Existing systems struggle with static models in rapidly changing conditions.

Purpose of the Study:

• To develop reconfigurable analog in-memory computing hardware for efficient reinforcement learning.
• To demonstrate the capabilities of split-gate MoS2 memtransistors for adaptive AI.
• To improve the performance of AI agents in real-time decision-making tasks.

Main Methods:

• Fabrication of split-gate molybdenum disulfide (MoS2) memtransistors.
• Utilizing local field-effect gating for precise control of memristive switching.
• Benchmarking adaptive reinforcement learning on a cartpole balancing task.

Main Results:

• Achieved improved control over memristive switching ratios and conductance states.
• Demonstrated efficient reinforcement learning with nonvolatile synaptic weight updates and rapid parameter adjustments.
• Showcased a 6-fold increase in total reward and a 5-fold reduction in programming steps for the cartpole task.

Conclusions:

• Split-gate MoS2 memtransistors offer a promising hardware platform for real-time, on-chip learning.
• The developed devices significantly enhance adaptive reinforcement learning capabilities for edge AI.
• This advancement paves the way for more sophisticated and responsive AI agents in robotics and autonomous systems.