Функціональна стійкість програмно-апаратної реалізації нейронної мережі: NEUROCOMP

Software system for ensuring the functional stability of neural networks  at the level of individual neurons. Unlike classic Dropout, the system provides resistance to hardware failures during network operation. Includes redundant training modules (Dropout, Adam, mini-batch), neuronal importance analysis, dynamic network reconfiguration, and results visualization. The system maintains accuracy in case of failure of up to 15% of neurons.Implemented in C# using the .NET Framework, Newtonsoft.Json for serialization, and ScottPlot for visualization. Supports saving/loading model state, interactive neuronal shutdown, and automatic architecture thinning.

Manager / participants: professor Bychkov O.S.