An FSCV Deep Neural Network: Development, Pruning, and Acceleration on an FPGA.

Area of Science:

• Neuroscience
• Electrochemistry
• Computer Engineering

Background:

• Fast-scan cyclic voltammetry (FSCV) measures brain neurotransmitter dynamics.
• Current FSCV data analysis is computationally intensive and not real-time capable.
• Low-resource devices require efficient, automated analysis methods.

Purpose of the Study:

• To develop a hardware-software co-design for real-time FSCV data analysis.
• To implement a pruned deep neural network (DNN) on a low-resource FPGA platform.
• To enable portable and efficient neurotransmitter concentration prediction.

Main Methods:

• Developed a DNN for dopamine concentration prediction and electrode identification.
• Pruned the DNN to reduce computational complexity.
• Implemented the pruned DNN using a custom overlay on an FPGA (PYNQ-Z2).

Main Results:

• Achieved 97.2% recognition accuracy with a 3.18 compression ratio for the pruned DNN.
• The FPGA implementation achieved a 13 ms execution time.
• Demonstrated low power consumption of 1.479 W on the PYNQ-Z2 board.

Conclusions:

• The hardware-software co-design enables real-time FSCV data analysis.
• Portable FPGA platforms can effectively host pruned DNNs for electrochemical sensing.
• This approach facilitates on-site neurotransmitter monitoring in neuroscience research.