Neural Decision Tree for Bio-TinyML

Deep neural networks(NN) are widely popular candidates for introducing machine learning in resource-constrained devices(tinyML). However, NNs in tinyML applications come with constraints such as high memory footprint and power consumption. Another popular candidate is decision trees (DT), which are lightweight, but in DT - choosing the tree depth, the best feature, and the threshold for subsetting data is not end-to-end as in parameter tuning of NNs. Hence, we here design a Neural Decision Tree (NDT), which can be trained using backpropagation in an end-to-end fashion and is simultaneously lightweight. We evaluate NDT for potential healthcare applications in Epileptic Seizure, Cervical Cancer, Breast Cancer, Parkinson’s, and Heart Disease Detection. Careful studies reveal that NDT performs at par with NNs but is computationally simpler with fewer neurons and with a hugely reduced number of multiply-accumulate operations(MACs). In addition, we propose methods for interpreting NDT behavior. Moreover, we perform systematic ablation studies on NDT to justify the design choices.