MFM Speedup | Fang Tian

This was a research internship project at Prof. B.T. Thomas Yeo’s lab at the National University of Singapore. The goal of this project is to leverage machine learning and deep learning to accelerate the parameter optimization process for the mean-field model(MFM), or more precisely parametric MFM (pMFM).

The pMFM is a computational model of human brains that simulates brain activities through an ordinary differential equation (ODE) system to produce realistic functional MRI (fMRI) signals. The model is used to study the dynamics of brain activity and to understand how different brain regions communicate with each other.

Illustration of the pMFM, for more details, please refer to the original paper

The original pMFM’s parameter optimization process utilizes Euler integration to numerically solve the ODE system, which can be computationally expensive to run. Therefore, I aim to speed up the parameter optimization process by using machine learning techniques to predict the MFM parameters’ performances without the need to solve ODEs. Specifically, I performed the following tasks:

Generated a synthetic dataset of pMFM parameters and their corresponding performance.
Validated two assumptions.
Explored different machine learning models to predict the pMFM parameters’ performances.
Explored various optimization techniques such as Covariance matrix adaptation evolution strategy (CMA-ES) and Gradient Descent.

In the end, I have demonstrated the feasibility of using machine learning to accelerate the pMFM parameter optimization process, achieving a speedup of 5,000x compared to the original Euler integration method. This project has led to a paper (Zeng* et al., 2024) that we are currently preparing for submission to Nature Methods. For more details of this project, please visit the GitHub repository.

References

2024

Optimizing Biophysically-Plausible Large-Scale Circuit Models With Deep Neural Networks

Tianchu Zeng^*, Fang Tian^* , Shaoshi Zhang, Gustavo Deco, Theodore D. Satterthwaite, Avram Holmes, and B.T. Thomas Yeo

2024

Abs PDF

Biophysically plausible large-circuit models provide valuable mechanistic insights into neural dynamics. However, setting up a model involves numerous free parameters, which are often manually set, leading to assumptions that may not align with biological plausibility. In contrast, optimizing these parameters results in more biologically plausible models. Previous papers have shown that spatially heterogeneous parameters improve the model performance in generating realistic neural dynamics. Prior works have employed iterative forward simulations via Euler integration to optimize the parameters, which require solving ordinary differential equations (ODE) for all regions simultaneously and each time step sequentially. This process poses a computational hurdle for an efficient model inversion. To tackle this problem, we propose DELSSOME (Deep Learning for Surrogate Statistics Optimization in Mean Field Modeling), which integrates deep learning models into neural mass model framework to circumvent the need to solve ODEs. These deep learning models are designed to learn surrogate statistics, which directly provide outputs required for optimization, from the input local circuit parameters. Our experiments demonstrate that our approach leads to a remarkable 500,000% speed-up of parameter optimization process compared to the traditional approach while achieving comparable performances. Moreover, our approach successfully replicates previous findings on excitation-inhibition ratio changes related to neurodevelopment and cognition. Overall, our results suggest the efficiency and broader applicability of the proposed approach in uncovering intricate neural dynamics.