Emotion and music are intrinsically connected, and researchers have had limited success in employing computationalmodels to predict perceived emotion in music. Here, we use computational dimension reduction techniques to discovermeaningful representations of music. For static emotion prediction, i.e., predicting one valence/arousal value for each 45smusical excerpt, we explore the use of triplet neural networks for discovering a representation that differentiates emotionsmore effectively. This reduced representation is then used in a classification model, which outperforms the original modeltrained on raw audio. For dynamic emotion prediction, i.e., predicting one valence/arousal value every 500ms, we examinehow meaningful representations can be learned through a variational autoencoder (a state-of-the-art architecture effectivein untangling information-rich structures in noisy signals). Although vastly reduced in dimensionality, our model achievesstate-of-the-art performance for emotion prediction accuracy. This approach enables us to identify which features underlieemotion content in music.

The ability to extract meaningful relationships from sequences is crucial to many aspects of perception and cognition, such as speech and music. This paper explores how leading computational techniques may be used to model how humans learn abstract musical relationships, namely, tonality and octave equivalence. Rather than hard-coding musical rules, this model uses an unsupervised learning approach to glean tonal relationships from a musical corpus. We develop and test a novel input representation technique, using a perceptually-inspired harmonics-based representation, to bootstrap the model’s learning of tonal structure. The results are compared with behavioral data from listeners’ performance on a standard music perception task: the model effectively encodes tonal relationships from musical data, simulating expert performance on the listening task. Lastly, the results are contrasted with previous findings from a computational model that uses a more simple symbolic input representation of pitch.

Statistical Learning (SL), the ability to extract probabilistic information from the environment, is a subject of much debate.It appears intuitive that such a profound mechanism of learning should carry predictive power towards general cognitiveability. Yet, previous attempts have struggled to link SL ability to measures of general cognitive function, suffering fromlow correlations and mediocre test-retest reliability. Here, we deploy a new continuous auditory SL task that achieves hightest-retest reliability ( r = .8) and shows that SL ability does significantly correlate with general cognitive function (up tor =. 56). Results are discussed in light of i) the theoretical implications of the high test-retest reliability of our novel SLtask, ii) SL ability as a marker of general cognitive function, and iii) future methodological considerations.