Ternary Neural Network in C#

The author explores the promising field of neural network quantization, moving from standard FP32 formats to ternary logic. The article examines the evolution of model training formats and provides a practical implementation of a Microsoft Bitnet 1.58 architecture equivalent using C#. Special attention is paid to extreme weight optimization, which enabled the creation of a model weighing just 29 KB, or 17.1 KB in its most compressed form. The article offers a deep technical analysis of how ternary logic can drastically reduce memory and computational requirements while maintaining neural network functionality. The author analyzes the potential of such compact models within the context of future AI development trends and invites readers to explore the project's source code to understand the principles of ternary neural networks.