An Oscillator Ensemble Model of Sequence Learning.

Alexander Maye,Peng Wang,Jonathan Daume,Xiaolin Hu,Andreas K Engel

doi:10.3389/fnint.2019.00043

Abstract

Learning and memorizing sequences of events is an important function of the human brain and the basis for forming expectations and making predictions. Learning is facilitated by repeating a sequence several times, causing rhythmic appearance of the individual sequence elements. This observation invites to consider the resulting multitude of rhythms as a spectral “fingerprint” which characterizes the respective sequence. Here we explore the implications of this perspective by developing a neurobiologically plausible computational model which captures this “fingerprint” by attuning an ensemble of neural oscillators. In our model, this attuning process is based on a number of oscillatory phenomena that have been observed in electrophysiological recordings of brain activity like synchronization, phase locking, and reset as well as cross-frequency coupling. We compare the learning properties of the model with behavioral results from a study in human participants and observe good agreement of the errors for different levels of complexity of the sequence to be memorized. Finally, we suggest an extension of the model for processing sequences that extend over several sensory modalities.

Highlights

Oscillations are a ubiquitous phenomenon when brain activity is observed at a sufficiently high temporal resolution, e.g., using EEG/MEG, or invasive methods
Great progress has been made toward understanding the functional role of oscillations in cognitive processes (Singer, 1999; Engel et al, 2001, 2013; Canolty and Knight, 2010; Giraud and Poeppel, 2012; Fries, 2015)
Their rhythmic nature suggests that neuronal oscillations could be used by the brain for learning, recognizing and producing rhythmic patterns in the interaction with the environment, and corresponding mechanisms have been suggested and studied in computational models

Summary

Introduction

Oscillations are a ubiquitous phenomenon when brain activity is observed at a sufficiently high temporal resolution, e.g., using EEG/MEG (electro-/magneto-encephalography), or invasive methods. Great progress has been made toward understanding the functional role of oscillations in cognitive processes (Singer, 1999; Engel et al, 2001, 2013; Canolty and Knight, 2010; Giraud and Poeppel, 2012; Fries, 2015) Their rhythmic nature suggests that neuronal oscillations could be used by the brain for learning, recognizing and producing rhythmic patterns in the interaction with the environment, and corresponding mechanisms have been suggested and studied in computational models. Oscillator-based models have replicated many of the properties of human memory for serial order (Brown et al, 2000) To this end, the two most relevant computational mechanisms are the encoding of arbitrary time intervals by an ensemble of oscillators with different periods and the dynamic adjustment of oscillation frequency and phase. This dynamic context has a number of desirable properties for learning sequences of events: First, despite the cyclic activity of the individual oscillators, the vector of the combined phases repeats over very

Methods

Results

Conclusion