FPGA-BASED IMPLEMENTATION OF A SYNTHESIZABLE WEIGHT PRUNING ALGORITHM FOR DEEP NEURAL NETWORKS: AN EXPERIMENTAL APPROACH
Artificial Intelligence (AI), using deep learning techniques led to intensive computational resource consumption due to deep network with many layers and large parameters. Pruning is one compression technique that can combat these issues by reducing unnecessary weights while maintain model accuracy. In this paper, a method is proposed to pruning small weights in fully connected layers, using comparator circuits to reduce unnecessary connections in the neural network. The model was deployed on a highly parallel device platform Field Programmable Gate Array )FPGA(, the experimental result on twenty pruning configurations have shown that pruning design achieves a utilization improvement up to 6.56%, frequency enhancement of 13.8% and a power consumption reduction of 37.9% over the baseline design. The output value remains at 0.98 for the experiments without any retraining phase, thus demonstrating the effectiveness of the proposed strategy in achieving high performance and energy efficiency without incurring loss in accuracy
Authors
- Esam Alkaldy (ORCID: https://orcid.org/0000-0002-8014-7488)
- Zainab A. Kadhim
Publication Details
- Journal
- Kufa Journal of Engineering
- Published
- 2026-11-14
- DOI
- https://doi.org/10.30572/2018/kje/170436
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00