Machine learning and deep learning—A review for ecologists

Abstract

The popularity of machine learning (ML), deep learning (DL) and artificial intelligence (AI) has risen sharply in recent years. Despite this spike in popularity, the inner workings of ML and DL algorithms are often perceived as opaque, and their relationship to classical data analysis tools remains debated. Although it is often assumed that ML and DL excel primarily at making predictions, ML and DL can also be used for analytical tasks traditionally addressed with statistical models. Moreover, most recent discussions and reviews on ML focus mainly on DL, failing to synthesise the wealth of ML algorithms with different advantages and general principles. Here, we provide a comprehensive overview of the field of ML and DL, starting by summarizing its historical developments, existing algorithm families, differences to traditional statistical tools, and universal ML principles. We then discuss why and when ML and DL models excel at prediction tasks and where they could offer alternatives to traditional statistical methods for inference, highlighting current and emerging applications for ecological problems. Finally, we summarize emerging trends such as scientific and causal ML, explainable AI, and responsible AI that may significantly impact ecological data analysis in the future. We conclude that ML and DL are powerful new tools for predictive modelling and data analysis. The superior performance of ML and DL algorithms compared to statistical models can be explained by their higher flexibility and automatic data‐dependent complexity optimization. However, their use for causal inference is still disputed as the focus of ML and DL methods on predictions creates challenges for the interpretation of these models. Nevertheless, we expect ML and DL to become an indispensable tool in ecology and evolution, comparable to other traditional statistical tools.

A Comparison of Deep Learning Architectures for Inferring Parameters of Diversification Models from Extant Phylogenies

“…From theory, one would expect that lower total error of DL methods can be explained by the bias-variance trade-off, which allows DL methods to trade off bias against a lower total error ( e.g. Pichler & Hartig, 2023). Looking at the decomposition of the error, however, we do not find that the deep-learning methods show greater bias than the MLE ( e.g.…”

“…This is in line with general practice in machine learning. To achieve a lower error, machine learning methods often trade off bias against variance (Pichler and Hartig 2023), whereas in statistics, bias is often seen as more crucial than variance, because an unbiased estimator allows a field to accumulate evidence over time (Shmueli 2010). Yet, given that there are usually no independent replicates of a phylogeny, we find it defensible to use the total error as the primary performance metric.…”

“…When considering the choice of network architecture (Pichler and Hartig 2023), the simplest option would be a standard fully connected Deep Neural Network (DNN). The disadvantage of this option is that fully connected DNNs do not perform favorably with high-dimensional structured input data (such as phylogenies).…”

Realising the promise of large data and complex models

“…In general, there is a trade-off between the complexity of the model being fitted and the associated intricacy of the information that can be extracted (given suitable and available data). Furthermore, statistical learning (or machine learning) techniques are rapidly increasing in their prominence and usage within ecology (Ho & Goethals, 2022;Pichler & Hartig, 2022), with such techniques often demonstrating good predictive performance, but at the lack of ecologically interpretable parameters. It is becoming increasing important to extract interpretable and meaningful results/output from appropriate models fitted to real data, combined with intelligent visualisations, within and beyond the wider scientific community, for example, with policy-makers One particular area of ecology in which increasing model complexity leads to further interpretability challenges is that of species' distribution modelling.…”

“…Such approaches are likely to have an important role in the future direction of methods in the ecological domain (Pichler & Hartig, 2022), particularly when prediction is a primary objective. However, such methods should not simply be blindly applied to align with popular analytical trends—it is important that there is a methodological driver underpinning their usage.…”