Self-supervised method for accelerated training of neural communication receivers

A method for training and using a neural communication receiver, the method comprising:

• receiving, through a radio frequency (RF) communications channel, RF signals carrying messages;
• digitizing the received RF signals to produce message signals;
• pretraining an encoder model by using contrastive self-supervised learning based on a first set of message signals to form respective representation signals of the message signals, wherein the first set includes positive pairs of message signals associated with similar pretraining messages and negative pairs of message signals associated with dissimilar pretraining messages, wherein representation signals of message signals in a positive pair are more similar than representation signals of message signals in a negative pair;
• training a neural receiver model by using supervised learning based on a second set of message signals to reconstruct training messages associated with the second set of message signals from representation signals formed by the encoder model from the second set of message signals;
• obtaining a message signal associated with a test message;
• applying the encoder model to the obtained message signal to form a representation signal;
• applying the neural receiver model to the representation signal to reconstruct the test message associated with the obtained message signal;
• and
• outputting the reconstructed test message.

A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• pretraining an encoder model by using contrastive self-supervised learning based on a first set of message signals to form respective representation signals of the message signals, wherein the first set includes positive pairs of message signals associated with similar pretraining messages and negative pairs of message signals associated with dissimilar pretraining messages;
• training a neural receiver model by using supervised learning based on a second set of message signals, wherein the second set of message signals are processed by the encoder model to produce representation signals that are provided as inputs to the neural receiver model to reconstruct corresponding training messages;
• and
• applying the encoder model and the neural receiver model to a message signal associated with a test message to reconstruct the test message.