An Operator Theoretic Approach for Analyzing Sequence Neural Networks

Analyzing the inner mechanisms of deep neural networks is a fundamental task in machine learning. Existing work provides limited analysis or it depends on local theories, such as fixed-point analysis. In contrast, we propose to analyze trained using an operator theoretic approach which rooted Koopman theory, Analysis Neural Networks (KANN). Key our method operator, linear object that globally represents dominant behavior network dynamics. The linearity facilitates via its eigenvectors and eigenvalues. Our reveals latter eigendecomposition holds semantic information related workings. For instance, highlight positive negative n-grams sentiments task; similarly, capture salient features healthy heart beat signals ECG classification problem.