Deep learning as a tool for ecology and evolution

Marek Borowiec, Paul B. Frandsen, Rebecca B. Dikow, Alexander J. McKeeken, Gabriele Valentini, Alexander E. White2021 · DOI

TL;DR

This review introduces deep learning (DL) as a powerful new tool for ecology and evolution, highlighting its ability to uncover complex patterns in large datasets. It explores DL's diverse applications, from species identification to population genetics, and provides a practical guide for biologists looking to incorporate it into their research.

What Problem Does This Paper Address?

Ecology and evolutionary biology are increasingly data-rich fields, generating vast amounts of information from sources like genomic sequencing, camera traps, and environmental sensors. Traditional analytical methods, while powerful, are often computationally expensive and struggle to handle the complexity and scale of this modern data. This creates a need for more efficient and effective tools to extract meaningful insights. Deep learning, with its capacity to learn intricate patterns directly from raw data, offers a promising solution to these challenges.

Methods

The paper provides a gentle introduction to deep learning, focusing on artificial neural networks (ANNs). It explains how ANNs, inspired by the brain, learn by adjusting connections between nodes (neurons) through processes like backpropagation and gradient descent. The review categorizes DL approaches by their application:

  • Supervised Learning: Used when ground truth data is available (e.g., labeled images of species). This involves training a network on a dataset and then testing its performance on unseen data.
  • Unsupervised Learning: Applied when no ground truth is available, for tasks like data visualization and clustering. The goal here is to reconstruct the original input after a transformation.
  • Reinforcement Learning: Used for optimizing actions in an environment, with feedback in the form of rewards.

The paper also discusses various neural network architectures, including Convolutional Neural Networks (CNNs) for image recognition, Recurrent Neural Networks (RNNs) for sequential data, and Variational Autoencoders (VAEs) for dimensionality reduction and data generation.

Datasets

The paper does not focus on specific datasets for training or evaluation but rather discusses the types of data DL can handle. These include:

  • Images (herbarium specimens, camera trap photos)
  • Audio (bird songs, bat sounds)
  • Time-series data (environmental monitoring, population dynamics)
  • Genomic data (DNA sequences, alignments)
  • Text data

The authors emphasize that DL's strength lies in its ability to process diverse data types, often by converting them into numerical tensors.

Key Insights and Ideas

  • DL as a Pattern Recognizer: Deep learning excels at identifying complex, hierarchical patterns that are difficult or impossible for humans or traditional algorithms to discern.
  • Beyond Likelihood-Based Methods: DL offers an alternative to traditional likelihood-based inference, particularly when models are complex or unknown.
  • "Black Box" Potential: While DL models can be highly accurate, their internal workings can be opaque, posing challenges for hypothesis testing and understanding the biological basis of predictions.
  • Data is King (and Can Be Biased): The performance of DL models is heavily dependent on the quality and quantity of training data. Biases in data collection can lead to biased model outputs, a critical consideration in ecological and evolutionary research.
  • Transfer Learning is Powerful: Pre-training models on large, general datasets (like ImageNet) and then fine-tuning them for specific ecological tasks can significantly improve performance and reduce the need for massive custom datasets.

Key Results

The paper is a review and primer, not a research study with specific quantitative results. However, it highlights the broad success of DL across numerous ecological and evolutionary applications:

  • Automated Species Identification: DL, particularly CNNs, has achieved high accuracy in identifying species from images, audio, and even DNA sequences, spanning a vast range of taxa.
  • Environmental Monitoring: DL is being used to monitor populations, detect diseases in crops, and analyze environmental data from sensors.
  • Behavioral Studies: DL enables non-invasive tracking of animal movement, posture, and behavior classification, revolutionizing how animal behavior is studied.
  • Genomics, Population Genetics, and Phylogenetics: DL is being applied to tasks like DNA sequencing, variant calling, inferring population structure, and even reconstructing evolutionary trees.

Limitations and Future Work

  • Data Quality and Bias: The paper repeatedly stresses the sensitivity of DL to biased or insufficient training data. Future work needs to focus on developing methods to mitigate these biases and ensure data representativeness.
  • Interpretability ("Black Box" Problem): Understanding why a DL model makes a certain prediction remains a challenge. Future research should focus on developing more interpretable DL models and techniques for dissecting their decision-making processes.
  • Computational Resources: While DL is becoming more accessible, training complex models still requires significant computational power, often necessitating the use of GPUs.
  • Ethical Considerations: As DL is applied to sensitive data, ethical implications, particularly regarding data bias and its impact on outcomes, need careful consideration.

The authors envision DL becoming an integral part of the future biologist's toolkit, driving advancements in areas like automated discovery, "open world learning" (discovering novel data classes), and more realistic simulations of biological processes.

Summary updated

Abstract

Deep learning is driving recent advances behind many everyday technologies, including those relying on speech and image recognition, natural language processing, and autonomous driving. It is also gaining popularity in biology, where it has been used for automated species identification, environmental monitoring, behavioral studies, DNA sequencing, and population genetics and phylogenetics, among other applications. Deep learning relies on artificial neural networks for predictive modeling and excels at recognizing complex patterns. Operating within the machine learning paradigm, deep learning can be viewed as an alternative to likelihood-based inference methods. It has desirable properties, including good performance and scaling with increasing complexity, while posing unique challenges such as sensitivity to bias in input data. In this review we provide a gentle introduction to deep learning, review its applications in ecology and evolution, and discuss its limitations and efforts to overcome them. We also provide a practical primer for biologists interested in including deep learning in their toolkit and identify its possible future applications.
BibTeX
@misc{borowiec2021deep,
  author = {Marek Borowiec and Paul B. Frandsen and Rebecca B. Dikow and Alexander J. McKeeken and Gabriele Valentini and Alexander E. White},
  title = {{Deep learning as a tool for ecology and evolution}},
  year = {2021},
  doi = {10.32942/osf.io/nt3as},
}
Figure