Resource-constrained FPGA/DNN co-design.

Area of Science:

• * Computational Neuroscience
• * Electrical Engineering
• * Analytical Chemistry

Background:

• * Deep neural networks (DNNs) excel in learning tasks but require substantial computational resources.
• * Executing complex DNNs on low-resource platforms necessitates hardware-software co-design strategies.
• * Electrochemical methods are vital for neurotransmitter concentration analysis.

Purpose of the Study:

• * To develop a miniaturized deep neural network (DNN) for analyzing electrochemical data.
• * To reduce DNN resource utilization for deployment on low-resource processing platforms.
• * To enable efficient neurotransmitter concentration determination using DNNs on FPGAs.

Main Methods:

• * A DNN model was applied to analyze electrochemical data for neurotransmitter concentration.
• * A DNN miniaturization algorithm combined pruning (weight reduction) and Lempel-Ziv-Welch compression.
• * A DNN overlay integrated decompression and inference for FPGA execution on a PYNQ-Z2 board.

Main Results:

• * The DNN miniaturization algorithm achieved a compression factor of 6.18.
• * Resource utilization on the FPGA was reduced by approximately 50%.
• * The approach successfully executed the DNN without complex quantization algorithms.

Conclusions:

• * Hardware-software co-design effectively reduces DNN complexity for low-resource applications.
• * Miniaturized DNNs are suitable for real-time analysis of electrochemical data, such as neurotransmitter concentrations.
• * The proposed method offers a practical solution for deploying DNNs on FPGAs, enhancing efficiency and reducing hardware requirements.