A Deep Learning Approach leveraging Transformer style Architectures for signal classification
DeepWave is a experimental Deep learning framework which utilizes recently developed Transformer based network architectures, such as OpenAI GPT-2(https://github.com/openai/gpt-2) and the BERT(https://github.com/google-research/bert) network, to create a new angle to classify spectrogram based signals.
Due to the time-series based nature of singal analysis and vector based sequence indices of the [time x frequency] based spectrogram an opportunity to utilize large and highly researched NLP based networks was apparant. DeepWave adapts the NLP transformer networks which utilize the convept of attention to provide a robust network infrastructure to classify signals.
Intial research is being conducted using th RadioML(https://www.deepsig.io/datasets) dataset which includes 220,000 artificially generated Radio Wave modulations with labels for their types. This dataset has been used as a Benchmark for many signal processing apporaches which will provide a baseline for DeepWave benchmarking.
Stay tuned for more!