MAtt: A Manifold Attention Network for EEG Decoding

Authors: Yue-Ting Pan, Jing-Lun Chou, Chun-Shu Wei

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The evaluation of the proposed MAtt on both time-synchronous and -asyncronous EEG datasets suggests its superiority over other leading DL methods for general EEG decoding.
Researcher Affiliation Academia Yue-Ting Pan Jing-Lun Chou Chun-Shu Wei National Yang Ming Chiao Tung University, Hsunchu, Taiwan wei@nycu.edu.tw
Pseudocode Yes Algorithm 1 Manifold attention module Require: A sequence of SPD data { xi}m i=1 , transformation matrices Wq, Wk, Wv 1: for i = 1: m do 2: qi = Wq xi W T q ; ki = Wk xi W T k ; vi = Wv xi W T v 3: end for 4: i, j {1, , m}, A := [αij]m m = 1 1+log(1+δL(qi,kj)) 5: A = Softmax(A) 6: for i = 1: m do 7: v i = Exp l=1 α il Log(vl) 8: end for 9: return a sequence of SPD data {v i}m i=1
Open Source Code Yes Source codes are available at https://github.com/CECNL/MAtt.
Open Datasets Yes We incorporate the BCI Competition IV 2a Dataset (BCIC-IV-2a) [47] to assess the performance on time-asynchronous motor-imagery (MI) EEG decoding , the MAMEM EEG SSVEP Dataset II (MAMEM-SSVEP-II) [48] and the BCI challenge error-related negativity (ERN) dataset (BCI-ERN) [49] to assess the performance on time-synchronous SSVEP and ERN EEG decoding.
Dataset Splits Yes For the BCIC-IV-2a dataset, we used the first session of a subject to the training set where one out of eight was used for validation for MAtt with m = 3. The model with the lowest validation loss within 350 iterations was used for testing on the second session of the same subject. For the MAMEM-SSVEP-II/BCI-ERN dataset, we assigned the first four sessions of a subject to the training set where one out of four was used for validation for MAtt with m = 7/m = 3. The model with the lowest validation loss within 180/130 iterations was used for testing on the fifth session of the same subject.
Hardware Specification No The paper describes the EEG recording hardware used for data collection (e.g., EEG electrodes, EGI 300 Geodesic EEG System) but does not provide specific details about the computational hardware (e.g., GPU models, CPU types, or cloud instances) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers needed to replicate the experiment.
Experiment Setup Yes Our parameter setting follows [19]. [...] We set ϵ as 1e-5 in our source code. [...] For the BCIC-IV-2a dataset, we used the first session of a subject to the training set where one out of eight was used for validation for MAtt with m = 3. The model with the lowest validation loss within 350 iterations was used for testing on the second session of the same subject. For the MAMEM-SSVEP-II/BCI-ERN dataset, we assigned the first four sessions of a subject to the training set where one out of four was used for validation for MAtt with m = 7/m = 3. The model with the lowest validation loss within 180/130 iterations was used for testing on the fifth session of the same subject.