New accurate and flexible design procedure for a stable KWTA continuous time network

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks

Background:

• The classical continuous time recurrent (Hopfield) network is a foundational model in neural computation.
• Adapting such networks for specific computational tasks like selection is crucial for advancing machine intelligence.

Purpose of the Study:

• To adapt the Hopfield network for K-winner-take-all operation, enabling it to identify and rank the K largest elements in a given list.
• To ensure a stable and reliable binary separation of outputs, distinguishing between the K largest and the rest.

Main Methods:

• Utilized sigmoidal neurons with controllable gain (G) and amplitude (m), interconnected by conductance (p).
• Implemented a resetting procedure for initial conditions and determined bias current (M) constraints for stable steady-state separation.
• Employed high gain and exploited sigmoid properties and network symmetry for asymptotic stability.

Main Results:

• The adapted network successfully performs a matching dynamic process between list element order and output order.
• Achieved a binary steady-state separation of outputs, with the top K outputs exceeding a threshold (+ξ) and others falling below (-ξ).
• Developed a synthesis procedure for the tanh sigmoid, computing parameter bounds (p, G, ξ, θ, M) from given network specifications (N, K, I, z(min), m).

Conclusions:

• The adapted Hopfield network effectively signals the ranks of the K largest elements in a list.
• The proposed synthesis procedure is compatible with network constraints and offers flexibility.
• Numerical tests provided insights into the method's qualities and limitations.