On-Chip Hardware Training: Implementation of Backpropagation for Parameter Fine-Tuning of Deep Neural Networks
Akash Dev Roshan, Prithwijit Guha, Gaurav Trivedi · 2025
On-chip training is a promising approach for enabling efficient machine learning in localized environments, offering advantages such as reduced dependence on cloud infrastructure, support for remote deployment, and autonomy from external networks. However, implementing backpropagation directly on hardware remains challenging due to its high computational complexity and resource demands. This paper proposes a novel, resource-efficient method for on-chip training that leverages a freeze-layer technique inspired by transfer learning. By freezing the initial layers of the model and finetuning only the final layers, the approach significantly reduces hardware overhead while enabling real-time adaptation of model parameters. Experimental results demonstrate that the proposed design achieves lower hardware utilization compared to existing state-of-the-art methods, making it well-suited for real-time deep neural network applications.