Neural Network Bias in Analysis of Galaxy Photometry Data

Advancements in the ability to collect, store, and access astronomical data have made a major impact on astronomy research in the past two decades. These changes reinforced the need for methodology that can analyze large and complex astronomical databases. One of the common tools used to approach that task is machine learning (ML), and specifically artificial neural networks (ANN). One of the primary downsides of ANNs is that they follow complex and non-intuitive data-driven rules, making it virtually impossible to formally describe the way they analyze the data. Here we analyze possible systematic biases that can affect the results observed when applying ANNs to photometry data. The results show that ANNs can lead to systematic biases in the annotation of the data. These biases are difficult to detect and profile, and can behave in a non-intuitive manner. Therefore, catalogs and data products annotated by neural networks should be used with extra caution.