Loading…
CNS*2020 Online has ended
Welcome to the Sched instance for CNS*2020 Online! Please read the instruction document on detailed information on CNS*2020.
Back To Schedule
Monday, July 20 • 9:00pm - 10:00pm
P91: A unified framework for the application and evaluation of different methods for neural parameter optimization

Log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Máté Mohácsi, Sára Sáray, Márk Patrik Török, Szabolcs Káli

Virtual room: https://meet.google.com/nyz-mzvw-cpm

Software: https://github.com/KaliLab/optimizer

Currently available experimental data make it possible to create complex multicompartmental conductance-based models of neurons. In principle, such models can approximate the behavior of real neurons very well. However, these models have many parameters and some of these parameters often cannot be directly determined in experiments. Therefore, a common approach is to tune parameter values to bring the physiological behavior of the model as close as possible to the experimental data. Rather than tuning the parameters by hand, a more principled way of determining good model parameters is to carry out a systematic parameter search using an appropriate global optimization algorithm. Although many such algorithms have been developed and applied successfully in various domains, and high-quality general implementations of many popular algorithms are available, the majority of these solutions have not been tested in a neural context. Our goal in this study was to create a software tool that provides uniform access to a large variety of different optimization algorithms; to develop a set of benchmark problems for neural parameter tuning; and to systematically evaluate and compare the various algorithms and implementations using our software and benchmarking suite.
We have created an updated and enhanced version of our previously developed software tool. In Optimizer, model evaluations can be performed either by the NEURON simulator (handled internally) or any external (black-box) simulator. All functionalities can be accessed from the graphical user interface; there is also a command line interface for batch processing. The new version was developed in Python 3 to support recent open-source Python modules. The repertoire of algorithms was extended by several new methods that proved effective in other studies. For many of these search algorithms, parallel optimization is also supported. A wide variety of features (including those in the eFEL package) can be used to evaluate the error of the optimization; multiple, weighted features are also supported. Our optimization tool currently supports about fifteen different optimization algorithms implemented by four separate Python packages: Inspyred, Pygmo, BluePyOpt, and Scipy.
Our neural optimization benchmark suite includes six separate problems that differ in complexity, model type, simulation protocol, fitness functions, and the number of unknown parameters. Our examples range from the classical Hodgkin-Huxley model (3 conductance parameters) to an extended integrate-and- fire model (10 parameters) and a morphologically and biophysically detailed hippocampal pyramidal cell (16 parameters). Some of our benchmarks use target data generated by a neuronal model with known parameters. However, in most of our benchmarks, the target data were recorded in physiological experiments, or were generated by more complex models than the one we were fitting.
We then tested the various algorithms on the different model optimization tasks, and compared the final error (after 10,000 model evaluations) and also the convergence speed (Fig. 1). We found that several evolutionary and related search algorithms delivered consistently good results across our entire test suite, even for higher-dimensional, multi-objective problems. Therefore, we would recommend trying these algorithms first for novel optimization problems. We also hope to extend our test suite with new problems and algorithms.

Speakers
avatar for Szabolcs Kali

Szabolcs Kali

senior research scientist, Institute of Experimental Medicine
Our research group uses various mathematical and simulation tools to study the dynamics and functions of both single neurons and networks in the hippocampus, often in combination with experiments conducted in the lab. Some of the main focus areas of our group are the following... Read More →



Monday July 20, 2020 9:00pm - 10:00pm CEST
Slot 10