Performance of Deep and Shallow Neural Networks, the Universal Approximation Theorem, Activity Cliffs, and QSAR

摘要: Neural networks have generated valuable Quantitative Structure-Activity/Property Relationships (QSAR/QSPR) models for a wide variety of small molecules and materials properties. They grown in sophistication many their initial problems been overcome by modern mathematical techniques. QSAR studies almost always used so-called "shallow" neural which there is single hidden layer between the input output layers. Recently, new potentially paradigm-shifting type network based on Deep Learning has appeared. learning methods impressive improvements image voice recognition, are now being applied to modelling. This paper describes differences approach deep shallow networks, compares abilities predict properties test sets 15 large drug data (the kaggle set), discusses results terms Universal Approximation theorem how DNN may ameliorate or remove troublesome "activity cliffs" sets.