Robustness Analysis on Dual Neural Network-based $k$ WTA With Input Noise.

Area of Science:

• Computational Neuroscience
• Machine Learning Theory
• Artificial Neural Networks

Background:

• The Dual Neural Network-based WTA (DNN-WTA) model is a key component in various computational tasks.
• Understanding the impact of input noise is crucial for the reliability and robustness of neural network models.
• Previous research has not fully characterized the behavior of DNN-WTA models under specific noise conditions.

Purpose of the Study:

• To investigate the effects of uniform and Gaussian input noise on the DNN-WTA model's convergence and output accuracy.
• To develop a formula for verifying the correctness of the network's outputs.
• To establish theoretical bounds on the probability of correct output generation.

Main Methods:

• Analysis of network state convergence under uniform and Gaussian input noise.
• Derivation of a formula to assess output correctness.
• Development of lower bounds for correct output probability with uniformly distributed inputs.
• Derivation of conditions for correct output generation based on minimum input separation.

Main Results:

• The DNN-WTA model's state converges to an equilibrium point under both uniform and Gaussian input noise.
• A formula is derived to determine if the network produces correct outputs.
• Lower bounds on the probability of correct outputs are established for uniformly distributed inputs.
• Conditions for correct output generation are identified when minimum input separation is known.

Conclusions:

• The study provides a theoretical framework for understanding DNN-WTA model behavior under input noise.
• The derived formula and probability bounds offer practical tools for assessing network reliability.
• The findings are applicable to scenarios involving random drift in comparators, enhancing practical relevance.