Who Makes the Choice? Artificial Neural Networks in Science and Non-Uniqueness

Machine learning techniques have become an essential part of many scientific inquiries, promoting novel discoveries. Here we distinguish between the output-oriented approach which regards neural networks as black boxes, and the feature-oriented approach which seeks to reveal and scientifically adopt the features captured by the network for the purpose of exploration and novelty. Focusing on the latter, we point at an issue of non-uniqueness when choosing between three types of features – mathematical, diagnostic, and real-world features. Scientists make choices among numerous features and rationalize their choices with background assumptions, but when aiming at exploration in an immature domain, rationalization neither justifies the choice nor guarantees that the chosen features are real. As a result, we propose that machine-captured features for this purpose should not be used as full-fledged evidence, but scientists should focus on the instrumental value of these features, such as refining existing descriptions or methods and inspiring future directions of research. We also suggest promoting the transparency of feature selection rationale and the plurality of choices.