Data-driven Tapered Spectral Density Features from Time Series Data

Music City Center 

We develop a parameterized and learnable statistical model of the frequency domain characteristics of signals, inspired by tapered spectral density methods. The proposed model is designed to capture intricate time-frequency dependencies. We show how we can learn the model from data using a contrastive training approach, and how the proposed model can encompass the classical multi-taper spectral density type models. We validate our model on two types of time series: optogenetically-evoked neural recordings and acoustic scene audio recordings. We evaluate the effectiveness of the learned representations on a classical in-domain classification task, and a cross-domain task to explore its external validity. Preliminary experimental results show the promises and the challenges of the proposed approach.

signal processing

feature representation

self-supervised learning

optogenetics

time series

multitaper 

Section on Statistical Learning and Data Science