How machine learning is revolutionising animal behaviour studies

Scientists explore the advantages of machine learning in studying animal behaviour

By Orson Auvermann

Recent technological advancements allow large datasets to be collected on the movement, fine-scale motion, social interactions, vocalizations and physiological responses of individual animals.[1] However, collecting replicated data, especially of wild animal populations, faces logistical difficulties and often leads to small sample sizes with up to thousands of factors to consider. Our increasing efficiency in collecting big data is therefore currently bottlenecked by our analytic abilities.

A new study has aimed to introduce machine learning in the field of animal behaviour research to make sense of these voluminous data sets. Three case studies were carried out in conducting this research.

In the first of these studies, an unsupervised learning algorithm was used to assign pheasant eggs to their respective nests (as female pheasants tend to lay eggs communally). The algorithm reduced egg variance to a limited set of meaningful features, by using expansive spectral and morphological data, and then assigned the eggs to their respective mothers. This tracking approach can be used to observe the difference in pheasant behaviour when brooding related or unrelated eggs, without the need for intrusive genetic sampling or direct observation.

The second case study showed jackdaws being fitted with PITs (passive integrated transponders), with a PIT antenna situated in a feeder. Social networks were mapped by extracting their foraging events from a continuous data stream. With previous research on jackdaw populations being limited to enclosures, this unsupervised experiment revealed information on the relationships between unrelated individuals. Mapping of this social data can lead to a better understanding of avian social learning and epidemiology.

In the final case study, aerial photos taken of the Serengeti National Park were used, with 64 x 64 pixel segments being manually labelled as either wildebeest, zebra, trees, rocks or grass. This information was fed to a learner for automatic classification. Currently, these photos have to be classified by hand, which is incredibly time consuming. Controlling the wildebeest migration is essential in monitoring the state of the ecosystem.

In the first two case studies, `deep learning’ structures were used to make inferences based on raw data, without the implementation of hand-crafted features. Information was analysed and processed through the learner and, through iterative analysis and backpropogation, a prediction could be made. In the wildebeest study, deep convolutional nets could be implemented to better process images, but initial classification may still be required as the training data set.

Machine learning eliminates the need for theory and allows large data sets to be modelled with minimal human error. The success of these studies grants an insight into the potential of machine learning to enable better understanding and prediction of animal behaviour and suggests that it will become an essential analytical toolkit.

References

  1. J. J. Valletta, C. Torney, M. King, A. Thornton and J. Madden, Applications of machine learning in animal behaviour studies, Anim. Behav. 124, pp. 203–220.

Books