News

Machine learning as a tool for the natural sciences: Opportunities and challenges
Giovanni Volpe
Invited Talk at BRC Day “Biomaterials meets AI”, University of Gothenburg, Gothenburg, Sweden, 12 November 2019

Abstract: Data-driven machine-learning methods are more and more widely used in the natural sciences. Machine learning offers unprecedented opportunities, but it also poses unexpected practical and fundamental challenges. Most importantly, machine-learning methods often work as black boxes, and therefore it can be difficult to understand and interpret their results. Here, we present an overview of the current state of the art of the adoption of machine learning in active-matter research. Finally, we discuss the opportunities and challenges that are emerging, highlighting how active matter and machine learning can work together for mutual benefit.

Bio: Giovanni Volpe is Associate Professor at the University of Gothenburg, where he leads the Soft Matter Lab (http://www.softmatterlab.org/).
He has published more than 80 articles on diverse topics including optical trapping, active matter, neurosciences, and machine learning.
He has co-authored the book “Optical Tweezers: Principles and Applications” (Cambridge University Press, 2015).
He is the recipient the ERC Starting Grant ComplexSwimmers, coordinator of the MSCA Innovative Training Networks ActiveMatter, and the KAW research grant “Active Matter Goes Smart”.
He is one of the chairs of the Conference Emerging Topics in Artificial Intelligence at the SPIE Optics & Photonics Meeting held annually in San Diego (CA).

Machine learning for active matter
Giovanni Volpe
Introductory Talk at CECAM Workshop “Active Matter and Artificial Intelligence”
CECAM-HQ-EPFL, Lausanne, Switzerland
30 September – 2 October, 2019

Data-driven machine-learning methods are more and more widely used in the natural sciences. Active-matter research is no exception and has recently started experiment- ing machine-learning approaches. Machine learning offers unprecedented opportunities, but it also poses unexpected practical and fundamental challenges. Most importantly, machine-learning methods often work as black boxes, and therefore it can be difficult to understand and interpret their results. Here, we present an overview of the current state of the art of the adoption of machine learning in active-matter research. Finally, we discuss the opportunities and challenges that are emerging, highlighting how active matter and machine learning can work together for mutual benefit.

Deep Learning Applications in Digital Video Microscopy and Optical Micromanipulation
Saga Helgadottir, Aykut Argun, Giovanni Volpe
Invited talk at RIAO/Optilas 2019, Cancun, Mexico, 23-27 September 2019

Since its introduction in the mid 90s, digital video microscopy has become a staple for the analysis of data in optical trapping and optical manipulation experiments [1]. Current methods are able to predict the location of the center of a particle in ideal condition with high accuracy. However, these methods fail as the signal-to-noise ratio (SNR) of the images decreases or if there are non-uniform distortions present in the images. Both these conditions are commonly encountered in experiments. In addition, all these methods require considerable user input in terms of analysis parameters, which introduces user bias. In order to automatize the tracking process algorithms using deep learning have been successfully introduced but have not proved to be usable for practical applications.

Here, we provide a fully automated deep learning tracking algorithm with sub-pixel precision in localizing single particle and multiple particles’ positions from image data [2]. We have developed a convolutional neural network that is pre-trained on simulated single particle images in varying conditions of, for example, particle intensity, image contrast and SNR.

We test the pre-trained network on an optically trapped particle both in ideal condition and challenged condition with low SNR and non-uniform distortions [3]. This pre-trained network accurately predicts the location the trapped particle and a comparison of detected trajectories, the distribution of the particle position and the power spectral density of the particle trajectory clearly shows that our algorithm outperforms tracking by radial symmetry [4]. Our algorithm is also able to track non-ideal images with multiple Brownian particles as well as swimming bacteria that are problematic for traditional methods.

In conclusion, our algorithm outperforms current methods in precision and speed of tracking non-ideal images, while eliminating the need for user supervision and therefore the introduction of user biases. 

References

[1] John C Crocker, David G Grier, Journal of Colloid and Interface Science 179, 298–310 (1996).

[2] Saga Helgadottir, Aykut Argun, Giovanni Volpe, Optica 6, 506–513 (2019).

[3] Philip H Jones, Onofrio M Maragò, Giovanni Volpe, Optical tweezers: Principles and applications. Cambridge University Press, 2015.

[4] Raghuveer Parthasarathy. Nature Methods 9724 (2012).

Soft Matter Meets Deep Learning
Giovanni Volpe
Department of Chemistry, University of Gothenburg, Sweden
19 September 2019

After a brief overview of artificial intelligence, machine learning and deep learning, I will present a series of recent works in which we have employed deep learning for applications in photonics and active matter. In particular, I will explain how we employed deep learning to enhance digital video microscopy [1], to estimate the properties of anomalous diffusion [2], and to improve the calculation of optical forces. Finally, I will provide an outlook for the application of deep learning in photonics and active matter.

References

[1] S. Helgadottir, A. Argun and G. Volpe, Digital video microscopy enhanced by deep learning. Optica 6(4), 506—513 (2019)
doi: 10.1364/OPTICA.6.000506

[2] S. Bo, F Schmidt, R Eichborn and G. Volpe, Measurement of Anomalous Diffusion Using Recurrent Neural Networks. arXiv: 1905.02038