Background
The objective of this research was to replace traditional, power-hungry deep neural networks with edge-capable, spiking neural networks to quantify cognitive load based on biological signals. Currently, deep learning methods that are used in classifying cognitive load from biological signals have not been implemented into real-time psychological monitoring due to resource-constrained devices. A possible remedy for this resource constraint lies in the field of neuromorphic computing. Neuromorphic computing is based on neural network simulation. It applies neural network principles to computer systems and algorithm design to achieve high-performance and low-power consumption for intelligent information processing applications. One algorithmic subset of neuromorphic computing is a spiking neural network (SNN) and is considered the third generation of artificial neural networks (ANNs). SNNs more closely model the behavior of a living nervous system as they consider both spatial and temporal aspects of input data for building the computational model. SNNs, when coupled along with neuromorphic hardware, provide an avenue to reduce energy consumption that allows real-time intelligence on the edge. This potentially minimizes the round-trip delay in decision making, lowers communication costs, and enhances security with local algorithms to process the data.
Approach
We used an open-source dataset that measures physiological signals to infer the neurological status of individuals using biomarkers from a wrist-worn wearable. Labels include physical stress, cognitive stress, emotional stress, and relaxation of 20 healthy subjects. The data was collected using non-invasive, wrist-worn biosensors and consists of electrodermal activity, temperature, acceleration, heart rate, and arterial oxygen level. The dataset was designed with relaxation periods between periods of physical, cognitive, and emotional stressors. We first developed a data preprocessing pipeline to align, standardize, and sample the data. For preprocessing, we attenuated the cognitive load from a value of 0 to 1 and calculated the best threshold to binarize the predictions. To calculate the accuracy, we optimized the threshold of the validation set and used the threshold that had the highest accuracy on the validation set onto the test set. Next, we compared a recurrent neural network (RNN) with a leaky integrate-and-fire (LIF) SNN in terms of accuracy and computational efficiency in classifying cognitive load in real-time. We simulated computational energy consumption of the networks on a CPU and Loihi 2 neuromorphic hardware.
Accomplishments
We calculated the distributions of metrics on the held-out test set, which consisted of five subjects with an area under the curve (AUC) of the receiving operator characteristic (ROC) to quantify model performance. Figure 1 shows the ROC curve of the SNN and the RNN. In addition, we calculated the accuracy and the F1 scores based upon the best preforming hyperparameter configuration of the SNN and RNN. The accuracy of the SNN was approximately 94.6%, whereas the accuracy of the RNN was approximately 84.9%. Both models outperformed the state-of-the-art accuracy of 84.6% for this dataset. In addition to predictive performance, we also measured the power consumption and inference time as additional metrics for success of this project. While we anticipated a 15% reduction in inference time and power consumption for a successful implementation, we found that the SNN used 21.6% less time and 27.4% less power for inferences compared to the RNN.
Figure 1: ROC curve with AUC comparison between RNN and SNN.