In his dissertation, M.Sc. (tech) Mikko Lehtimäki shows that the studied method is effective for reducing the computational burden of brain simulations. The method retains more information than existing acceleration methods, and is effective for speeding up different classes of mathematical models, ranging from synaptic plasticity to deep artificial neural networks.
Simulation studies are used in neuroscience to understand how different molecules, cells and networks in the brain work together. With this understanding, we can design better therapies and medication for brain disorders. The knowledge can also be used to further develop artificial intelligence.
"Currently, simulation studies need to make a difficult choice - are you going to create a very detailed model of a tiny fraction of a neuronal system, or are you going for
a brain-scale, yet abstract representation of brain anatomy and activity? Even with the best supercomputers, you cannot have both", says Lehtimäki.
Insights towards new applications of deep artificial neural networks
In order to create large-scale brain simulations, researchers use very crude descriptions of neuronal cells. Such simplified models do not always allow us to see deep into the mechanisms or biology of neural phenomena. Simple models are a very useful tool, but they have limitations. The methods proposed in Mikko Lehtimäki's dissertation address this question.
There are principled methods to derive approximations of large-scale mathematical models – especially if the systems are linear, meaning that interactions follow a certain set of rules. The models used in neuroscience do not typically fit in this category. For this reason, the present dissertation is focused on the complex, nonlinear models that are common in neuroscience.
“We use model order reduction techniques, which aim to derive compressed representations of nonlinear dynamics. What’s very interesting is that although we would simulate a compressed model, we can still reconstruct an approximation of all the details of the original model”, Lehtimäki explains.
The dissertation provides results on a wide range of models, used in different types of simulations. With model order reduction methods, researchers could add more information or scale to their simulation studies, improving the insights gained from mathematical modeling in neuroscience.
The dissertation also shows that these methods can be used in the field of machine learning, to accelerate a class of deep learning models. Very deep artificial neural networks are computationally demanding to use, and often not suitable for real-time applications on consumer hardware. The results of the dissertation are hence a step towards bringing deep neural networks into new applications.
The research towards this dissertation was a joint effort between the Computational Neuroscience Group, Faculty of Medicine and Health Technology at Tampere University, headed by Docent, Principal investigator Marja-Leena Linne, and the Systems Theory Research Group, Faculty of Information Technology and Communication Sciences at Tampere University, headed by Associate Professor Lassi Paunonen. The EU FET flagship Human Brain Project (HBP), where the Computational Neuroscience Group from Tampere participated, provided an exciting and educative network to discuss the results of this dissertation and its implications in neuroscience. One example is the integration of the studied methods to neuronal simulation tools. The neuroscience infrastructure development started in the HBP now continues in the EBRAINS network in Europe.
Mikko Lehtimäki has an M.Sc. in computational biology and signal processing. He is a co-founder of Softlandia, a solution-driven deep-tech software consultancy, where Lehtimäki leads data and machine learning projects.
Public defence on Friday 16 June
The doctoral dissertation of M.Sc. (Tech.) Mikko Lehtimäki in the field of biomedical engineering, Model Order Reduction for Modeling the Brain, will be publicly examined at the Faculty of Medicine and Health Technology of Tampere University at 13 o’clock on Friday June 16, 2023 in the auditorium FA032 of the Festia building, (Korkeakoulunkatu 8, Tampere). Dr. Jennifer S. Goldman from the Ernst Strüngmann Institute for Neuroscience will be the opponent while Docent Marja-Leena Linne will act as the custos.
The dissertation is available online.
The event can be followed via remote connection (Panopto).
You can find Mikko Lehtimäki in Twitter: @mikkolehtimaki and LinkedIn.
Photo: Anni Koskela/Studio Kuje