Sunday, Aug 3: 4:00 PM - 5:50 PM
4028
Contributed Papers
Music City Center
Room: CC-103A
Main Sponsor
Section on Statistical Learning and Data Science
Presentations
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.
Keywords
signal processing
feature representation
self-supervised learning
optogenetics
time series
multitaper
The prediction performances of neural networks and the conventional inference method of seasonal time series analysis have been compared both computationally and theoretically. Theorems and extensive simulations have also been provided to demonstrate the superiority of Time Series Modelling over Neural Network Modelling.
Keywords
Activation Functions
Feed Forward Neural Network
Monte Carlo Simulation
Projection Pursuit Regression
Stochastic resonance (SR), a nonlinear phenomenon originally introduced in climate modeling, enhances signal detection by leveraging optimal noise levels within nonlinear systems. Traditional SR techniques, primarily based on single-threshold detectors, are limited to time-invariant signals and often require excessive noise for detecting weak signals, which can degrade complex signal characteristics. To address these limitations, this study explores multi-threshold systems and the application of SR in the frequency/multiscale domain using wavelet transforms. We propose a double-threshold detection system that integrates two single-threshold detectors to enhance weak signal detection. The proposed system is evaluated in both the original data and multiscale domains using simulated and real-world test signals, and its performance is benchmarked against existing detectors. Experimental results demonstrate that, in the original data domain, the proposed double-threshold detector significantly improves weak signal detection compared to conventional single-threshold approaches. Performance is further enhanced in the frequency domain, requiring lower noise levels.
Keywords
Stochastic resonance
Multiscale signal processing
Fisher information
Wavelet transforms
Predicting hidden states using time-series classification models can result in frequent and erratic switching between the predicted states. This is particularly evident in applications with a high temporal resolution and during transitions between states. In the current investigation, a hidden Markov model (HMM) is fitted to the predicted states from trained time-series classification models to smooth these predictions and eliminate any high-frequency and/or erratic state switching observed in the outcomes. The HMM smoothing approach used in this study is shown to be highly effective at this task and is demonstrated in a case study using both mini-rocket and long-short-term-memory time series state predictions.
Keywords
prediction smoothing
time series classification
hidden markov model
Addressing non-linear trends in longitudinal data with irregular measurements involves fitting linear trends in segments joined at fixed times, known as change points (CPs). Methods to determine CP locations and numbers for piecewise linear mixed effects models are scarce, standard software lacks adequate algorithms, and the RE-EM tree may emphasize the intercept over the slope of trends. The Segmented package in R is a powerful tool for analyzing segmented relationships and identifying changepoints in regression models, though it has limited use for longitudinal data. This study explores the application of segmented methods combined with grid search to accurately identify CPs in longitudinal data. The proposed two-step approach first estimates the number and initial locations of CPs without considering within-subject correlation. In the second step, these locations are refined by accounting for within-subject correlation through a grid search in a piecewise linear mixed effects model. The findings demonstrate that this combined method effectively optimizes CP locations and outperforms the RE-EM tree, in fitting non-linear early childhood growth patterns measured by BMIz.
Keywords
Change Point
Non-linear Curves
Longitudinal Data
RE-EM tree
piecewise linear mixed effects model
segmented methods
First Author
Md Jobayer Hossain, Nemours Biomedical Research, A.I. DuPont Children's Hospital
Presenting Author
Md Jobayer Hossain, Nemours Biomedical Research, A.I. DuPont Children's Hospital