Higher-order interactions, adaptivity, and phase transitions in a novel reservoir computing model

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Complex Systems

Background:

• Reservoir neural networks (RNNs) are powerful tools for time-series processing.
• Brain neural ensembles exhibit complex dynamics, including adaptivity and higher-order interactions.
• The 'edge-of-chaos' regime is hypothesized to optimize information processing in neural systems.

Purpose of the Study:

• To propose a novel reservoir neural network model integrating brain ensemble properties.
• To investigate the role of adaptivity, higher-order interactions, and phase transitions in computation.
• To evaluate the model's performance on benchmark machine learning tasks.

Main Methods:

• Developed a reservoir neural network incorporating adaptivity and higher-order interactions.
• Introduced a phase transition mechanism enabling 'edge-of-chaos' computations.
• Tested the network on tasks including multidimensional periodic pattern reproduction and Lorenz attractor prediction.

Main Results:

• Interelement couplings were identified as the primary drivers for generating target outputs.
• The network successfully reproduced complex time-series data.
• A novel phase transition was observed post-learning, altering network dynamics.

Conclusions:

• The proposed reservoir neural network effectively leverages brain-inspired principles for computation.
• The findings highlight the importance of interelement couplings and phase transitions in neural computation.
• The model's ability to undergo a post-learning phase transition suggests enhanced adaptability and computational capacity.