Using neural networks to create fast and reusable approximate likelihood functions for ACT-R

Likelihood functions form the basis for statistical inference techniques, including maximum likelihood estimation, and Bayesian estimation/model comparison. Unfortunately, deriving likelihood functions analytically for cognitive architectures such as ACT-R can be challenging, if not impossible in some cases, often requiring considerable time and expertise. Simulation-based approximations are computationally intensive, making them impractical to implement in real-time applications. We demonstrate how recently developed techniques for learning intractable likelihood functions with neural networks can be applied to a visual search model based on ACT-R, and reused once trained. Our work extends prior applications in two ways: (1) we demonstrate that the technique can be scaled to a large number of conditions based on the size of the visual search array, and (2) we demonstrate that the technique is applicable to both unimodal and multimodal versions of the model. We conclude with a discussion for scaling up neural network techniques for approximating likelihood functions.