[2511.05350] Perceptually Aligning Representations of Music by way of Noise-Augmented Autoencoders

View a PDF of the paper titled Perceptually Aligning Representations of Music by way of Noise-Augmented Autoencoders, by Mathias Rose Bjare and 4 different authors

View PDF
HTML (experimental)

Summary:We argue that coaching autoencoders to reconstruct inputs from noised variations of their encodings, when mixed with perceptually motivated losses, yields encodings which can be structured in keeping with a perceptual hierarchy. We show the emergence of this hierarchy by displaying that, after coaching an audio autoencoder on this method, perceptually salient info is captured in coarser illustration constructions than with standard coaching. Moreover, we present that such perceptual hierarchies enhance latent diffusion decoding within the context of estimating pitch surprisal in music and predicting EEG-brain responses to music listening. In each circumstances, our outcomes surpass these of earlier strategies. Pretrained weights can be found on this http URL.