Abstract
Advances in sensor technology and automation have ushered in an era of dataabundance, where the ability to identify and extract relevant information inreal time has become increasingly critical. Traditional filtering approaches,which depend on a priori knowledge, often struggle to adapt to dynamic orunanticipated data features. Machine learning offers a compellingalternative-particularly when training can occur directly at or near thedetector. This paper presents a digital hardware architecture designed forreal-time neural network training, specifically optimized for high-throughputdata ingestion. The design is described in an implementation-independentmanner, with detailed analysis of each architectural component and theirperformance implications. Through system parameterization, the study explorestrade-offs between processing speed, model complexity, and hardware resourceutilization. Practical examples illustrate how these parameters affectapplicability across various use cases. A proof-of-concept implementation on anFPGA demonstrates in-situ training, confirming that computational accuracy ispreserved relative to conventional software-based approaches. Moreover,resource estimates indicate that current-generation FPGAs can train networks ofapproximately 3,500 neurons per chip. The architecture is both scalable andadaptable, representing a significant advancement toward integrating learningdirectly within detector systems and enabling a new class of extreme-edge,real-time information processing.