Year: 2024
Type: article
Abstract: Figure 1: A tree of sampling functions and transformations as a new data structure and framework for synthetic data generation.Samples are generated by traversing the tree from the root to the leaves.Each path through the tree represents its own class-conditional distribution.The branching points represent categorical distributions which determine the path to take for a sample during the tree traversal.By changing the parameters of the transformations over time, such trees can represent evolving distributions suitable to generate data streams (see Figure 2). SummaryStreamGen is a framework for generating streams of labeled, synthetic data from trees composed of sampling functions and transformation monoids (see Figure 1).Due to the expensive nature of the labelling process, researchers and machine learning practitioners often rely on existing datasets and stochastic data augmentation pipelines like torchvision.transforms.Compose objects (TorchVision maintainers and contributors, 2016).While such methods and datasets are appropriate to study learning from static domains, emerging research fields like continual learning study learning on long streams of data, representing evolving experiences.StreamGen addresses this need by giving researchers a tool to model time-dependent, diverse class-conditional distributions.Such distributions can be represented through the use of a tree data structure (or other more general linked structures like directed acyclic graphs) to store sampling functions and Farthofer.(2024).StreamGen: a Python framework for generating streams of labeled data.
DOI: 10.21105/joss.07206
Author: Laurenz A. Farthofer
Institution: Graz University of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Award: G4406901660
OpenAlex ID: W4405110018