Part 2/8:
Before delving into the technical aspects of Totem, it is crucial to establish the core question driving this work: Can modern machine learning methods effectively model continuously evolving time series in a general manner across multiple tasks? This inquiry is at the heart of many contemporary data science challenges, motivating the need for methodologies that can jointly train on diversified data and maintain efficacy across various domains.
As time series data is inherently diverse — encompassing everything from weather data to neural signals — the tasks requiring attention include:
Imputation: Filling in corrupted signals.
Forecasting: Predicting future events based on past sequences.
Anomaly Detection: Identifying unusual patterns within the data.