Page 297 - Data Science Algorithms in a Week

278 Jose M. Prieto

INTRODUCTION

Artificial neural networks are a type of artificial intelligence method. They are
applied in many disparate areas of human endeavours, such as the prediction of stock
market fluctuations in economy, forecasting electricity load in energy industry,
production of milk in husbandry, quality and properties of ingredients and products in the
food industry, prediction of bioactivities in toxicology and pharmacology or the
optimization of separation processes in chemistry (Dohnal, Kuča & Jun, 2005; Goyal,
2013).
In particular, the prediction of the bioactivity of natural products after their unique
chemical composition is an idea already well established among the scientific community
but not systematically explored yet, due to the experimental complexity of characterising
all possible chemical interactions between dozens of components (Burt, 2004). In this
regard, Neural networks have an enormous advantage in that they do require less formal
statistical training, can detect complex non-linear relationships between dependent and
independent variables and all possible interaction without complicated equations, and can
use multiple training algorithms. Moreover, in terms of model specification, ANNs
require no knowledge of the internal mechanism of the processes but since they often
contain many weights that are estimated, they require large training set. The various
applications of ANNs can be summarized into classification or pattern recognition,
prediction and modeling (Agatonovic-Kustrin & Beresford, 2000; Cartwright, 2008).
Therefore, the use of ANNs may overcome these difficulties thus becoming a
convenient computational tool allowing the food and cosmetic industry to select herbal
extracts or essential oils with optimal preservative (antioxidant and antimicrobial
properties) or pharmacological activities (anti-inflammatory properties). This is not
trivial, as natural products are notoriously complex in terms of chemical composition,
which may significantly vary depending on the batch and the supplier. This variability
implies a constant use of laboratory analysis. ANNs able to model and predict such
properties would result in savings and enhanced consistency of the final product. The use
of such computational models holds potential to overcome –and take into account- all the
possible (bio) chemical interactions, synergisms and antagonisms between the numerous
components of active natural ingredients.

BASIC CONCEPTS IN ARTIFICIAL NEURAL NETWORKS

To facilitate the understanding of the non-specialised reader, this section is conceived
as a “layman” presentation of the fundamental concepts surrounding the use of ANNs.
For a deeper understanding, the reader is encouraged to read the excellent papers