Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Implementing Dynamic Causal Models in the Probabilistic Programming Framework With Applications to Brain Connectivity Modelling
Estimating causal relationships between parts of the brain is an ongoing challenge. The aim of this project is to transfer well-known models of brain connectivity to a probabilistic programming framework that is amenable to fast model iteration and flexible, composable inference methods.
Estimating causal relationships between parts of the brain is an ongoing challenge in the
neuroscience community. Dynamic Causal Models (DCM) have been shown to be highly effective in fitting models of directed brain connectivity. Due to the large amount of data and the intractability of exact inference in these models, variational methods are used for approximating the posterior.
However, there are well-known limitations on the power of variational methods for approximating Bayesian posteriors. In practice, traditional variational inference implementations require painstaking, manual computation of update steps that directly depend on the specific structure of each model. A small change in the model may require a completely new inference method. Thus, model iteration and triaging different inference
methods are often time-consuming.
Probabilistic programming aims at overcoming these limitations with a generalized modelling language that can be interpreted by an automated inference engine. While these methods are gaining wider adoption in the machine learning community, they are only beginning to gain adoption by the computational neuroscience community
For more details, see attached document.
Estimating causal relationships between parts of the brain is an ongoing challenge in the neuroscience community. Dynamic Causal Models (DCM) have been shown to be highly effective in fitting models of directed brain connectivity. Due to the large amount of data and the intractability of exact inference in these models, variational methods are used for approximating the posterior.
However, there are well-known limitations on the power of variational methods for approximating Bayesian posteriors. In practice, traditional variational inference implementations require painstaking, manual computation of update steps that directly depend on the specific structure of each model. A small change in the model may require a completely new inference method. Thus, model iteration and triaging different inference methods are often time-consuming.
Probabilistic programming aims at overcoming these limitations with a generalized modelling language that can be interpreted by an automated inference engine. While these methods are gaining wider adoption in the machine learning community, they are only beginning to gain adoption by the computational neuroscience community
For more details, see attached document.
- Become proficient in Pytorch and Pyro
- Become proficient in DCM-based models
- Apply DCMs to real-world fMRI data
- Implement regression DCM, which works in the frequency domain rather than
- time-domain in a probabilistic programming language.
- Interpret results
- Explore possible extensions for (r)DCM
- Become proficient in Pytorch and Pyro - Become proficient in DCM-based models - Apply DCMs to real-world fMRI data - Implement regression DCM, which works in the frequency domain rather than - time-domain in a probabilistic programming language. - Interpret results - Explore possible extensions for (r)DCM