Compensated Integrated Gradients for Reliable Explanation of Electroencephalogram Signal Classification

Figures

• 

  Figure 1. Diagram of the proposed method. (A) Preprocessing to compute the compensation amount (orange dashed arrow), given as the difference between the contribution to the classification obtained using the IG method with an arbitrary user-defined baseline (blue arrow) and the contribution to the classification obtained using the SS method that implicitly exhibits an appropriate baseline (green arrow). (B) Proposed IG method that compensates for the divergence from the appropriate baseline. This is equivalent to the path integral along the orange path.

• 

  Figure 2. Model architecture for EEG signal classification using the CNN with constraints. Time-series signals measured by electrode xi for i=1,…,n are fed into temporal convolution layers gj(xi) for j=1,…,m. The input features are not spatially convolved owing to the independent constraint. The processing gj(xi) consists of the same parameterization owing to the identical constraint. If a classifier satisfies these constraints, the compensated IG can reliably compute the spatial contributions of xi to the outputs.

• 

  Figure 3. Contributions of EEG electrodes over the scalp to classify alcoholism (A) and control (B) using the temporal CNNs. The average contributions of the methods applied to 100 data examples from the UCI EEG dataset are shown. The red and blue areas represent positive and negative contributions to the classifications, respectively. The proposed method involves the compensated integrated gradients (left panels), the Shapley sampling shows the desired contributions as a reference (middle panels), and integrated gradients show the contributions obtained using the original integrated gradients method with the zero-input baseline (right panels).

• 

  Figure 4. Contributions of EEG electrodes over the scalp to classify alcoholism (A) and control (B) using the spatiotemporal CNNs. The average contributions of the methods applied to 100 data examples from the UCI EEG dataset are shown. The red and blue areas represent positive and negative contributions to the classifications, respectively. The proposed compensated integrated gradients method is shown (left panels), while the Shapley sampling shows the desired contributions as a reference (middle panels), and integrated gradients show the contributions obtained using the original integrated gradients method with the zero-input baseline (right panels).