Online and Continual Learning of RF Data Using Brain-Inspired Architectures

Abstract

Developing systems capable of learning at the edge post-deployment can greatly expand AI’s ability to adapt to changes in the environment and incorporate new information that may not be available prior to its deployment. Biological systems have become extremely adept at processing and learning complex signals with few computational resources in real time. In this work, we use the insect’s processing and internal representation of polarized light and auditory signals as inspiration to demonstrate archictectures capable of learning new RF signals post-deployment. Our architecture works with the complex envelope as input, and creates a robust internal representation that projects the in-phase and quadrature components into a higher dimensional space. It also relies on modulated synaptic plasticity rules to learn specific RF patterns in real time. In order to validate our architecture we have constructed a dataset of words consisting of specific sequences of multiple symbols, and subject our architectures to two types of online learning assays: in the online learning task the architecture is receiving a randomized stream of RF signals and needs to learn the vocabulary in a supervised fashion in a single experiment in real time. In the continual learning task, different patterns are shown sequentially, and the system needs to demonstrate the ability to incorporate this new information and build a larger vocabulary while minimizing catastrophic forgetting. For this presentation we will focus on vocabularies encoded using 8-PSK and 16-QAM modulation schemes. However, the proposed approach is general enough that could be applicable to any type of signal amenable to extract its complex envelope. This research has been partially funded through DARPA’s Lifelong Learning Machine program and Argonne’s Laboratory Directed Research and Development program.