Dynamics brain-state allocation using echo state network
Poster No:
1682
Submission Type:
Abstract Submission
Authors:
Yolanda Vives-Gilabert,1, Felipe Torres2, Andre Gómez-Lombardi3, WAEL EL-DEREDY4
Institutions:
1Universitat de València, Valencia, Spain, 2Universidad de Valparaíso, Valparaíso, Valparaíso, 3Universidad de Valparaíso, Valparaíso, Chile, 4UNIVERSIDAD DE VALPARAISO, Valparaíso, Valparaíso
First Author:
Co-Author(s):
Introduction:
Transiently stable and recurrent patterns of activity in the spontaneous EEG are thought to represent fundamental computational properties of the brain shaping behavior and brain function, transitioning between various states or modes of activity based on the ongoing cognitive demands, environmental stimuli, and internal processes [Trujillo-Barreto et al, 2019]. The identification of the brain states and their transitioning is an area of active research [Woolrich et al 2013; Baker et al 2014 ;Honcamp et al 2022] . Of particular interest are brain state allocation methods that can handle the natural non-linear dynamics of the brain, in a manner that affords subsequent biological interpretability [Trujillo-Barreto et al, 2019]. Here we explore the potential of Echo State Network (ESN) modeling for dynamical brain state allocation based on neuroimaging data. ESN is a recurrent neural network with randomly connected neurons, and it acts as a nonlinear dynamic system, Jaeger (2002) and Lukoševičius (2012). The state of the reservoir captures the network's response to the temporal patterns and dynamics present in the input signals. When feeding EEG data into an ESN and train the network to predict the same EEG data at the output, the internal states of the ESN capture relevant temporal patterns, dependencies, and information present in the input EEG signals, reproducing its nonlinear dynamics. By using a compact representation inside the ESN [the reservoir], the states would encode the temporal dynamics of the EEG signals, and reflect how the EEG signals have evolved over time.
Methods:
An Echo State Network is a type of recurrent neural network with a specific architecture characterized by a fixed random reservoir of recurrently connected neurons, comprising three layers Jaeger (2002) and Lukoševičius (2012): An input layer that receives the incoming data [EEG] and communicates it to the reservoir. A reservoir layer comprising large number of recurrently connected neurons. The connections between neurons in the reservoir are randomly assigned and usually remain fixed during training [here we adapt the reservoir connections using Hebbian learning rule]. Finally, the Output (Readout) Layer) is the trainable part of the network usually using linear methods. We followed this convention to training the output layer of the ESN, with the output predicting the input, as in autoencoders, while training the internal reservoir connections using Hebbian learning (Yusoff,et al 2016). Hebbian learning adjusts the weights based on the (temporal) correlations between the activities of connected neurons. To test the method's capture of the non-linear dynamics, we simulated an EEG like time series with non-linear interaction between two frequencies [Figure 1a].
Results:
The unsupervised training of the reservoir connections, leads to robust internal state that represent the temporal dynamics of the time-series [Figure 1B]. The internal state trace a the evolution of the oscillatory modes that emerge due to the frequency interactions.
Conclusions:
Echo State Networks offer a simple non-parametric machine learning model of nonlinear EEG type data that affords subsequent feature extraction, classification and regression analyses.
Modeling and Analysis Methods:
Classification and Predictive Modeling
EEG/MEG Modeling and Analysis 1
Methods Development 2
Keywords:
Machine Learning
Other - Reservoir computing
1|2Indicates the priority used for review
Provide references using author date format
Honcamp, H., Schwartze, M., Linden, D. E., El-Deredy, W., & Kotz, S. A. (2022). Uncovering hidden resting state dynamics: A new perspective on auditory verbal hallucinations. Neuroimage, 255, 119188.
Jaeger, H. (2002). Adaptive nonlinear system identification with echo state networks. Advances in neural information processing systems, 15.
Lukoševičius, M. (2012). A practical guide to applying echo state networks. In Neural Networks: Tricks of the Trade: Second Edition (pp. 659-686). Berlin, Heidelberg: Springer Berlin Heidelberg.
Trujillo-Barreto, N. J., Araya, D., & El-Deredy, W. (2019). The discrete logic of the Brain-Explicit modelling of Brain State durations in EEG and MEG. BioRxiv, 635300.
Woolrich, M. W., Baker, A., Luckhoo, H., Mohseni, H., Barnes, G., Brookes, M., & Rezek, I. (2013). Dynamic state allocation for MEG source reconstruction. Neuroimage, 77, 77-92.
Yusoff, M. H., Chrol-Cannon, J., & Jin, Y. (2016). Modeling neural plasticity in echo state networks for classification and regression. Information Sciences, 364, 184-196.