August 7, 2017
When machine learning meets neurobiology
Peter Rupprecht and Stephan Gerhard, two members of Rainer Friedrich's group at the FMI, have developed a powerful and elegant machine-learning approach to reconstruct action potentials from calcium signals in neurons. Their method won them the "Spikefinder challenge" - a friendly competition which attracted more than 30 teams of scientists - including deep-learning experts - from around the world.
Optical methods such as multiphoton microscopy are widely used to measure neuronal activity, but their readout is indirect: rather than directly detecting action potentials, they measure changes in the fluorescence of a calcium indicator, typically in neuronal somata. The fluorescence signals depend on calcium influx through voltage-gated calcium channels, which in turn are activated by action potentials. However, the precise relationship between calcium signals and action potentials remains ill-defined.
Various groups have therefore devised computer algorithms to reconstruct action potentials from calcium signals. To determine which method best describes this relationship, scientists at the Bernstein Center for Computational Neuroscience in Tübingen and the Howard Hughes Medical Institute (HHMI) Janelia Research Campus launched a competition - the "Spikefinder challenge".
Peter Rupprecht, a PhD student in Rainer Friedrich's group, thought that machine-learning approaches using multilayer convolutional neural networks - known as deep learning - should be ideal for tackling this problem. He developed an approach of this kind together with Stephan Gerhard, a postdoctoral fellow in Friedrich's lab who uses deep learning for image analysis. Rainer Friedrich explains: "Rather than using brute force, Peter and Stephan designed a relatively small network, preconfigured it using neurobiological know-how, and preprocessed the input data in a smart way." This original approach turned out to be very powerful: not only is it far superior to the approach previously used by Friedrich's lab, but it also won joint first prize in the Spikefinder challenge.
Key success factors were a good general understanding of deep-learning methods, a thorough knowledge of multivariate data analysis, and - most important, according to Friedrich - some very clever ideas. Rupprecht and Gerhard, along with the authors of the other top-performing algorithms in the challenge, have been invited to describe their method in a paper.
» Spikefinder challenge
» Spikefinder results
» Code on Github