Deep and Shallow Neural Networks

Artificial intelligence (AI) methods like neural networks have been used successfully for decades for QSAR and drug design and are now generating equally interesting models and predictions for materials more generally (see Applications Volume, Chapter 12). There has been much excitement and more than a little hype about the potential of neural networks to revolutionize many technologies, particularly when deep learning algorithms are employed. As with traditional neural networks, many researchers are interested in trying these potentially seductive new technologies simply because they are new and fashionable. There have been an increasing number of papers and books written on how deep learning methods will become paradigm shifting technologies in many areas, and drug discovery and chemoinformatics are not immune. Experienced practitioners of these research domains are also embracing deep learning, and papers are starting to emerge showing how effective the methods are. However, caution must be used to not allow deep learning to traverse yet another hype cycle, where new methods fail to live up to the high expectations imposed on them, inevitable disappointment results that damages the field, and a more mature appreciation of their value appears some years (or even decades in the case of neural networks) later. To master any method a deep understanding of the theory behind them and realistic expectations of their capabilities need to be developed. Neural networks are the closest practical thing to a universally applicable modeling tool, as their recent broad application in image processing, voice recognition, decision making, and quantitative modeling and prediction of properties of diverse materials has shown.