摘要：Artificial neural networks (ANNs) consist of multiple intermediate layers known as hidden layers stacked togetherthrough which the input is passed to obtain the desired output.The hidden layers are crucial for feature extraction which inturn impacts the performance of the entire model. However,they do not have a fixed number and the ideal value istraditionally derived iteratively by assessing the performanceof the architecture. It is thus desired to analyze and derivethe standard amount and deciding criteria for these hiddenlayers, as well as the impact of altering their configuration onthe performance of the system. In this paper, we present ourfindings for this problem when working on Spatio-temporaldata to predict the activity of drivers in semi-autonomousvehicles. The performance of the different architectures is assessed on the standard brain4cars dataset. Exploratory researchis conducted to understand the impact of variation in thehidden layers in a deep neural network architecture. A detailedmodeling procedure is followed out to present an unbiasedanalysis of the impact which the architectural changes holdon the performance of the system. The performance is assessedby considering the accuracy and other performance metricsof the system on the testing data. We also evaluate the timerequired by the system for delivering the inference. Both thesefactors are seen to be significantly affected by the architectureconfiguration.
关键词：Artificial neural networks; deep learning; LSTMs; RNNs; brain4cars
