Emotion is a complex psychological state that involves three distinct components, including subjective experiences, physiological responses and behavioral responses. As such, the human brain activity related to emotion would be highly variable and complicated, requiring sophisticated machine learning methods to infer emotional information from brain signals. This project harnesses human magnetoencephalography (MEG) to recognize personal emotional states from brain signals (Fig. 8A). Specifically, the project classifies MEG data into two classes, positive vs. negative valence. The MEG data were recorded while participants viewed short video clips edited to evoke various positive or negative emotions. A total of forty video clips were used in the experiment. The MEG features include spectral data over a range of frequency bands from 1 Hz to 95 Hz. Four machine learning models are compared, including FDA, SVM, Gaussian Process and MLP. The comparative analysis reveals that simpler classifiers with FDA and SVM outperforms Gaussian Process and MLP with accuracy around 85% (Fig. 8B). MoonA interprets this result in a way that MEG features can represent emotion-related information in a linearly separable space. But she also remarks that high-dimensional MEG data may need further processing to enable nonlinear classifiers to capture emotion-related information.