MasterThesis

Master thesis on continuous time representation in signal decoding tasks.

Abstract

In the signal decoding tasks, we work with multidimensional time series, which are a
discretization of a continuous process. The latest works in neural ODE illustrate the
possibility to work with recurrent neural networks as with differential equations.

This work addresses such applications as change of sampling rate and handling missed or
irregular data. It becomes possible if we represent our signal as a continuous in time
function. This approach is relevant for signals from various wearable devices:
accelerometers, heart rate monitors, devices for picking up brain signals such as
electroencephalograms or electrocorticograms.

The main result of this work is an algorithm which allows us to work with a signal as if
it was a continuous function. We also look at different applications of this algorithm and
propose to do further research on expanding the continuity of time to the continuity of
space.

Text

The full text of the thesis can be found
here

Experiments

For now it's all in the notebook, soon the training and visualizing would be done through
CLI (work in progress)