Modelling trial-by-trial changes in the mismatch negativity.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Neuroscience

Background:

• Mismatch negativity (MMN) reflects the brain's ability to learn environmental statistical structures and predict sensory inputs.
• The precise computational mechanisms and neurobiological basis of MMN remain debated.
• Existing theories offer diverse explanations for MMN generation.

Purpose of the Study:

• To introduce a mathematical framework for formalizing and testing computational theories of MMN.
• To compare the explanatory power of five major MMN theories against single-trial EEG data.
• To propose a unifying view of MMN based on approximate Bayesian inference.

Main Methods:

• Developed a mathematical framework to formalize MMN theories.
• Applied the framework to five distinct MMN theories, including predictive coding, model adjustment, and novelty detection.
• Tested model predictions against single-trial MMN amplitudes from an oddball experiment using EEG data.

Main Results:

• Three theories (predictive coding, model adjustment, novelty detection) were unified under the free-energy principle of approximate Bayesian inference.
• Models derived from the free-energy principle offered more plausible explanations for trial-by-trial MMN amplitude variations than traditional change detection and adaptation models.
• Empirical data supported the Bayesian learning interpretation of MMN.

Conclusions:

• MMN generation involves approximate Bayesian learning of sensory regularities.
• The MMN-generating process updates a probabilistic model of the environment based on prediction errors.
• The free-energy principle provides a unifying framework for understanding MMN.