DNN-kWTA With Bounded Random Offset Voltage Drifts in Threshold Logic Units.

Area of Science:

• Artificial Intelligence
• Neural Networks
• Analog Computing

Background:

• Dual neural network-based k-winner-take-all (DNN- k WTA) models identify the k largest inputs.
• Offset voltage drifts in threshold logic units (TLUs) can impact DNN- k WTA operational correctness.
• Existing research often assumes specific drift distributions, limiting applicability.

Purpose of the Study:

• To analyze the operational correctness of DNN- k WTA networks considering bounded TLU offset voltage drifts.
• To investigate both time-invariant and time-varying drift scenarios.
• To derive convergence conditions, operational probabilities, and convergence time metrics.

Main Methods:

• Theoretical analysis of DNN- k WTA network states under bounded TLU drifts ([-∆, ∆]).
• Derivation of sufficient conditions for correct network operation.
• Probabilistic analysis for uniformly distributed inputs and computation of convergence time.

Main Results:

• Convergence of the DNN- k WTA network state is proven for the time-invariant drift case.
• A lower bound for proper network operation probability is established as (1-2∆)ⁿ for uniformly distributed inputs.
• Methods for computing exact convergence time, mean, and variance are derived, applicable to both drift cases.

Conclusions:

• The study provides robust theoretical guarantees for DNN- k WTA network operation despite TLU drifts.
• Derived metrics offer insights into the network's stability and operational speed.
• Simulation experiments validate the theoretical findings, confirming the model's practical relevance.