Computational capabilities of random automata networks for reservoir computing

Area of Science:

• Computational neuroscience
• Complex systems dynamics
• Machine learning theory

Background:

• Reservoir computing (RC) leverages the dynamics of complex systems for computation.
• Traditional RC often uses homogeneous recurrent neural networks.
• Discrete, heterogeneous dynamical systems like Random Boolean Networks (RBNs) are less explored as reservoirs.

Purpose of the Study:

• To investigate the relationship between dynamics and computational capability in RBNs used as RC reservoirs.
• To test the hypothesis that intrinsic computation is maximal at the 'edge of chaos'.

Main Methods:

• Utilized Random Boolean Networks (RBNs) as discrete, heterogeneous reservoirs for reservoir computing.
• Extended RBNs with an input layer to allow external perturbation.
• Analyzed the trade-off between input separability and fading memory in RBN dynamics.

Main Results:

• Found that computational capability in RBN-based RC is linked to the balance between separability and memory.
• Optimal computational performance, specifically classification power, was observed at critical connectivity.
• Perturbing RBNs can prevent them from settling into attractors, enabling richer dynamics for computation.

Conclusions:

• Systems at the 'edge of chaos' exhibit maximal intrinsic computation, as demonstrated by RBNs in RC.
• Critical connectivity in RBNs provides an optimal balance for computational tasks.
• Results suggest potential for constructing novel computational devices using complex, heterogeneous systems like biomolecular substrates.