Recovering Neurocognitive Evidence Accumulation Models of Response Inhibition with Invertible Neural Networks

Response inhibition, the ability to withhold actions that are no longer appropriate, is commonly studied using the stop-signal task. Participants’ performance on the task is often understood as a race between a go- and stop-process, making it tempting to fit evidence accumulation models as the go- and stop-racers. Prior research shows, however, that behavioral stop-signal evidence accumulation models have unrecoverable parameters. In this study, we explored whether neurocognitive models of response inhibition are recoverable from neural and behavioral stop-signal data, as it has been suggested that the latency of the N200 peak reflects visual encoding time and marks the beginning of evidence accumulation. We wrote simulation-based neurocognitive models of stop-signal performance in Python, where a racing DDM and a Wiener process described the go- and stop-evidence accumulation processes, respectively. The simulated brain-behavioral measures along with the simulations themselves were used to train invertible neural networks using BayesFlow, that were then used for posterior inference. These trained networks recovered the neurocognitive models’ parameters and were moderately robust against high variance in the simulated EEG measures. We argue that racing Wiener processes can be used as measurement models for response inhibition and provide future researchers with a simply extendable framework for fitting their own neurocognitive models in BayesFlow, without having to derive any likelihood functions.