Recurrent neural network model of human event-related potentials in response to intensity oddball stimulation

Abstract

The mismatch negativity (MMN) component of the human event-related potential (ERP) is frequently interpreted as a sensory prediction-error signal. However, there is ambiguity concerning the neurophysiology underlying hypothetical prediction and prediction-error signalling components, and whether these can be dissociated from overlapping obligatory components of the ERP that are sensitive to physical properties of sounds. In the present study, a hierarchical recurrent neural network (RNN) was fitted to ERP data from 38 subjects. After training the model to reproduce ERP waveforms evoked by 80 dB standard and 70 dB deviant stimuli, it was used to simulate a response to 90 dB deviant stimuli. Internal states of the RNN effectively combine to generate synthetic ERPs, where individual hidden units are loosely analogous to population-level sources. Model behaviour was characterised using principal component analysis of stimulus condition, layer, and individual unit responses. Hidden units were categorised according to their temporal response fields, and statistically significant differences among stimulus conditions were observed for amplitudes of units peaking in the 0 to 75 ms (P50), 75 to 125 ms (N1), and 250 to 400 ms (N3) latency ranges, surprisingly not including the measurement window of MMN. The model demonstrated opposite polarity changes in MMN amplitude produced by falling (70 dB) and rising (90 dB) intensity deviant stimuli, consistent with loudness dependence of sensory ERP components. Although perhaps less parsimoniously, these observations could be interpreted within the context of predictive coding theory, as examples of negative and positive prediction errors, respectively.

Neural correlates of face perception modeled with a convolutional recurrent neural network

“…Recurrent neural networks (RNNs) solve this problem with recurrent connections that provide information from previous time-steps, thereby enabling the network to model sequences. These are widely used in natural language processing, and have recently been applied to model ERP waveforms [14][15][16][17] . Examples of RNN use in evoked response modeling include simulating ERPs to different stimuli 14,15 , estimating changes in ERP morphology between states of consciousness 17 , and exploring the computational processes of auditory-evoked potential generation [14][15][16] .…”

“…These are widely used in natural language processing, and have recently been applied to model ERP waveforms 14–17 . Examples of RNN use in evoked response modeling include simulating ERPs to different stimuli 14,15 , estimating changes in ERP morphology between states of consciousness 17 , and exploring the computational processes of auditory-evoked potential generation 14–16 .…”