Towards end-to-end likelihood-free inference with convolutional neural networks

Area of Science:

• Computational Statistics
• Machine Learning
• Bayesian Inference

Background:

• Complex simulator-based models necessitate advanced methods for parameter inference.
• Approximate Bayesian computation (ABC) is crucial for models with non-standard distributions.

Purpose of the Study:

• To introduce a novel, fast, and end-to-end approach for approximate Bayesian computation (ABC).
• To enable simultaneous derivation of posterior mean and variance directly from raw simulated data using neural networks.

Main Methods:

• Utilized fully convolutional neural networks for an end-to-end ABC approach.
• Trained neural networks on simulated data to map raw data to posterior moments.
• Developed reusable models applicable across different research contexts.

Main Results:

• The method successfully derived posterior means and variances for complex models.
• Demonstrated lower approximation error compared to existing machine learning ABC methods.
• Achieved performance comparable to probability density estimation (PDA) for the leaky competing accumulator (LCA) model.

Conclusions:

• The proposed fully convolutional neural network-based ABC method offers a significant advancement in parameter inference.
• This approach provides a fast, reusable, and accurate alternative to existing methods for complex statistical models.