ReFormer generates synthetic radio-frequency (RF) signals by compressing waveforms into a discrete codebook with a vector-quantized autoencoder and training a decoder-only transformer to autoregressively produce new token sequences that decode into realistic RF data. Students learn how discrete representation learning and autoregressive transformers combine to model signal distributions, how prompts condition the generated statistics, and why this is a scalable, tractable alternative to GAN- and diffusion-based RF synthesis for data augmentation.
Presents ReFormer, a generative model that synthesizes realistic radio-frequency (RF) signal samples to augment scarce RF datasets. RF waveforms are first compressed into sequences of discrete tokens…