For autonomous agents to operate effectively in dynamic environments, they need training algorithms that can continuously learn new tasks from streams of data in an online, task sequential and jointly optimal manner. This capability is especially important for aerospace applications that often deploy resource constrained platforms that lack the computer storage necessary for model training and/or the compute power necessary to rapidly digest new information. Furthermore, many aerospace missions require the ability to initialize a new model starting with very limited training data. This would allow a model to start learning directly from a small data stream and without requiring training data restrictions such as shuffled batches of stochastically mixed training classes.This paper introduces a new online, incremental learning algorithm called Ridge Regression for Rapid Class Augmentation (R3CA). It is based on a regularized version of the eXtending Rapid Class Augmentation (XRCA) algorithm that incrementally learns new classes without experiencing ‘catastrophic forgetting’ on its prior classes. R3CA was developed to improve XRCA’s performance in data-starved environments that are common for small model initialization and continuous learning on small streams of increasing information. The approach uses a ridge regression cost penalty to increase the solution’s numerical stability and minimize its overfitting for low sample support use cases. Results demonstrate that the new R3CA algorithm excels in low sample support while retaining XRCA’s ability to sequentially learn new classes while preserving performance on prior classes. Experiments show that neither the original XRCA algorithm nor other reduced-rank regularization methods can handle these low sample support scenarios as well. This makes R3CA an attractive option for continuous learning on remote platforms that need to learn on streaming data, in an online and real-time manner while having limited on-board memory and compute resources.
Ridge Regression for Rapid Class Augmentation
02.03.2024
2213257 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
IEEE | 1991
|ELM Ridge Regression Learning Algorithm of Ridge Parameter Optimization
British Library Online Contents | 2011
|Policy Capturing with Ridge Regression
Online Contents | 1996
|Ridge regression using prior information
Tema Archiv | 1974
|