Minimum Precision Requirements for Deep Learning with Biomedical Datasets

Deep neural networks (DNNs) are powerful machine learning models but are typically deployed in large computing clusters due to their high computational and parameter complexity. Many biomedical applications require embedded inference on resource-constrained platforms thus causing a challenge when considering the deployment of DNNs. One method to address this challenge is via reduced precision implementations. We use an analytical method to determine suitable minimum precision requirements of DNNs and show its application to the CHB-MIT EEG seizure detection dataset and the Bonn dataset for brain electrical activity recognition. We show that our method leads to 2 × reduction in average precision and 45% complexity reduction compared to the minimum uniform precision assignment. Compared to a conventional 16-b precision assignment, our method leads to 9 x complexity reduction. Furthermore, we study the impact of network topology on precision and accuracy. Once again we find our method to be 2× more efficient that the uniform assignment for all topologies considered.