August 17, 2017

Meet Peter Rupprecht

Peter Rupprecht is a PhD student in Rainer Friedrich's group studying computations in the olfactory system of adult zebrafish. Recently, he developed a powerful and elegant machine-learning approach to reconstruct action potentials from calcium signals in neurons. This method won him and Stephan Gerhard the "Spikefinder challenge" - a friendly competition which attracted more than 30 teams of scientists - including deep-learning experts - from around the world.

Q: What motivated you to participate in this challenge?
It is very natural for me to be interested in reliable inference of spikes from calcium data because I record this kind of data myself. During the last couple of years, a lot of papers on spike inference have been published, but it was never really possible to objectively compare algorithms among each other. The spikefinder challenge was therefore a big step forward, and that's why it made sense to me to invest some of my time here.
Apart from this, I was also keen on diving into the field of deep learning, a branch of machine learning that shares some central questions with neuroscience. However, deep learning unfolds its potential only for very large datasets. The spikefinder challenge was based on datasets that contained a large number of samples and were therefore ideally suited for deep learning. But the datasets were still of low complexity and therefore small enough to be processed on a personal computer without high-end GPUs. So I didn't think twice!

Q: You are no expert in deep-learning methods. How were you able to outcompete experts in the field?
There are deep learning packages out there (e.g. Keras) that are pretty easy to use for anybody who is not afraid of programming and parameter exploration; also, I had joined forces with Stephan Gerhard, who brought in his broader computer science knowledge and his programming skills. However, to optimize the deep net for our specific task, we did not only need expert knowledge about deep learning, but an equally deep understanding of the data that we wanted to analyze.
For the challenge, the dataset had been assembled from different subsets, each of which showed certain peculiarities, due to different calcium indicators, recording qualities or neuronal firing rates. After understanding these peculiarities in terms of descriptive statistics (which had nothing to do with deep learning), we could provide the algorithm with this distilled knowledge. This allowed the algorithm to generalize its predictions to unseen data and to improve its performance, and it is these improvements that made the difference.

Q: Can you apply these methods to your current projects?
Yes, I have already applied the algorithms to calcium imaging data that I had recorded before.

Q: You seem to have a knack for "thinking outside the box". You already developed a 3D multiphoton microscope based on bass speakers. How do you develop these ideas? And where do get your inspiration?
I guess that 'thinking outside of the box' is easier if one does not care about boxes. I believe that there are no real boundaries between science and life or between the different subfields of science. Which means that inspiration is basically everywhere, or at least wherever your understanding of things is deep enough. To give an example, if you do not think in fixed categories, it is easy to realize that both bass speakers for sound generation and scan devices for point scanning microscopes produce a high-pitched sound of similar frequency range (Hz-kHz), stemming in both cases from a mechanical part that moves really fast. The idea to use a loudspeaker-like device for a scanning microscope is therefore somewhat natural.
Developing the idea into something real is maybe the more difficult part, and it requires first of all a lot of hard work and a good intuition about which details and problems matter and which don't.
Finally, I think it is helpful to have somebody who likes your ideas or joins your project. For the microscope project, it was Rainer, my PhD supervisor, who supported my attempts. For the deep learning project, it was Stephan Gerhard who played this role. In both cases it was this external positive reinforcement that pushed the project to its final completion, and maybe both ideas would at some point have lost their momentum without this feedback from outside of my head.

» More about Peter's work
» More about the Spikefinder Challenge
» Where to download the code