Advances in imaging and recording throughput are generating neuroscience data at an ever-increasing rate, necessitating efficient data analysis approaches. This is particularly evident in subdisciplines such as connectomics, as well as the analysis of behavior or neuronal activity.

Image processing with machine-learning approaches. Reproduced in part from Dorkenwald, S. et al., Nat. Methods 14, 435–442, 2017. Credit: Katie Vicari/Springer Nature

Machine-learning and, in particular, deep-learning approaches can help process and analyze large volumes of data. In supervised deep learning, convolutional neural networks learn from training data by passing it through a multilayered network of simple modules. Such networks progressively abstract the data and extract features, which can be applied to analyze new data. Depending on the input data and the way the networks are trained, the extracted features can vary widely.

Image-based data, which are often analyzed manually, can benefit from advanced machine learning. For example, electron microscopic imaging of brain tissue routinely yields terabytes of data that are highly laborious to analyze and annotate manually. Tools such as the SyConn pipeline or the Multicut method (Nat. Methods 14, 435–442, 2017; Nat. Methods 14, 101–102, 2017) can efficiently segment, and, in the case of SyConn, further analyze such data sets.

Annotating and classifying animal behavior in data sets consisting of hours of video recordings is similarly tedious when conducted manually. Machine-learning tools such as classifiers as well as more advanced approaches have been applied to analyze, for example, Drosophila or mouse behaviors.

Analyzing calcium-imaging data is another area that has recently seen an influx of machine-learning methods. Such data are notoriously noisy, and extracting neuronal spikes is not trivial, especially since the recorded signal depends on the properties of the calcium indicator used. Supervised learning approaches are flexible enough that they can work on data sets from different calcium indicators than those they trained on.

In select examples, machine learning has proven to be a useful tool in the analysis of the growing deluge of data in neuroscientific research. However, the technology has yet to reach the mainstream. To be adopted by the broader community, machine-learning-based approaches will have to demonstrate their robustness under different conditions. And importantly, efforts will have to be made to facilitate their application and make them easy to use.