Learning communication systems using channel approximation

1. A method performed by at least one processor to train at least one machine-learning network to communicate over a communication channel, the method comprising:

obtaining first information;

using an encoder machine-learning network to process the first information and generate a first radio-frequency signal;

transmitting the first radio-frequency signal through a first communication channel;

determining a second radio-frequency signal that represents the first radio-frequency signal having been altered by transmission through the first communication channel;

simulating transmission of the first radio-frequency signal over a second communication channel implementing a channel machine-learning network, the second communication channel representing a model of the first communication channel;

determining a simulated radio-frequency signal that represents the first radio-frequency signal having been altered by simulated transmission through the second communication channel;

calculating a first measure of distance between the second radio-frequency signal and the simulated radio-frequency signal; and

updating the channel machine-learning network using the first measure of distance.

2. The method of claim 1 , further comprising:

using a decoder machine-learning network to process the simulated radio-frequency signal and generate second information that is a reconstruction of the first information;

calculating a second measure of distance between the second information and the first information; and

updating at least one of the encoder machine-learning network or the decoder machine-learning network using the second measure of distance.

3. The method of claim 2 , further comprising:

calculating a combined distance using the first measure of distance and the second measure of distance; and

updating at least one of the encoder machine-learning network, the decoder machine-learning network, or the channel machine-learning network using the combined distance.

4. The method of claim 2 , wherein the encoder machine-learning network, the decoder machine-learning network, and the channel machine-learning network of the approximated channel are updated concurrently.

5. The method of claim 2 , wherein the encoder machine-learning network or the decoder machine-learning network, and the channel machine-learning network are updated iteratively.

6. The method of claim 5 , wherein updating the channel machine-learning network comprises:

inserting a reference tone into the first radio-frequency signal at an offset time;

determining a time at which the first radio-frequency signal has passed through the first communication channel;

using the reference tone and the time at which the first radio-frequency signal has passed through the first communication channel to determine at least one of timing or frequency of the first radio-frequency signal;

receiving the second radio-frequency signal; and

using the timing or frequency of the first radio-frequency signal to remove alignment errors from the second radio-frequency signal.

7. The method of claim 5 , wherein updating the channel machine-learning network comprises:

using a digital to analog converter to process the first radio-frequency signal, wherein the digital to analog converter receives a signal from a time source;

using an analog to digital converter to produce the second radio-frequency signal, wherein the analog to digital converter receives a signal from the time source; and

using a known timing of the time source to align the second radio-frequency signal with the first radio-frequency signal.

8. The method of claim 2 , wherein at least one of the encoder machine-learning network, the decoder machine-learning network, or the channel machine-learning network of the approximated channel comprises one of a dense neural network (DNN), a convolutional neural network (CNN), or a recurrent neural network (RNN) comprising parametric multiplications, additions, and non-linearities.

9. The method of claim 2 , wherein the second measure of distance comprises at least one of (i) a cross-entropy between the second information and the first information, (ii) a geometric distance metric between the second information and the first information, or (iii) a measure of accuracy of one or more of the reconstructed bits, code words, or messages of the first information from the second information.

10. The method of claim 2 , wherein updating the encoder machine-learning network, comprises updating at least one encoding network weight or network connectivity in one or more layers of the encoder machine-learning network,

updating the decoder machine-learning network comprises updating at least one decoding network weight or network connectivity in one or more layers of the decoder machine-learning network, and

updating the channel machine-learning network comprises updating at least one network weight or network connectivity in one or more layers of the channel machine-learning network.

11. The method of claim 2 , further comprising:

processing the first radio-frequency signal to generate a first analog radio-frequency waveform that is input into the first communication channel;

receiving a second analog radio-frequency waveform as an output of the first communication channel that represents the first analog radio-frequency waveform having been altered by the first communication channel; and

processing the second analog radio-frequency waveform to generate the second radio-frequency signal.

12. The method of claim 1 , wherein the first measure of distance comprises at least one of (i) a cross-entropy between the second radio-frequency signal and the simulated radio-frequency signal, (ii) a geometric distance metric between the second radio-frequency signal and the simulated radio-frequency signal, (iii) a measure of probability distribution between second radio-frequency signal and the simulated radio frequency signal, or (iv) a measure of distance between characterizing the second radio-frequency signal and the simulated radio-frequency signal.

13. The method of claim 1 , wherein the first communication channel comprises at least one of a radio communication channel, an acoustic communication channel, or an optical communication channel.

14. A system comprising:

at least one processor; and

at least one computer memory coupled to the at least one processor having stored thereon instructions which, when executed by the at least one processor, cause the at least one processor to perform operations comprising:

obtaining first information;

using an encoder machine-learning network to process the first information and generate a first radio-frequency signal;

transmitting the first radio-frequency signal through a first communication channel;

determining a second radio-frequency signal that represents the first radio-frequency signal having been altered by transmission through the first communication channel;

simulating transmission of the first radio-frequency signal over a second communication channel implementing a channel machine-learning network, the second communication channel representing a model of the first communication channel;

determining a simulated radio-frequency signal that represents the first radio-frequency signal having been altered by simulated transmission through the second communication channel;

calculating a first measure of distance between the second radio-frequency signal and the simulated radio-frequency signal; and

updating the channel machine-learning network using the first measure of distance.

15. The system of claim of claim 14 , wherein the operations further comprise:

using a decoder machine-learning network to process the simulated radio-frequency signal and generate second information that is a reconstruction of the first information;

calculating a second measure of distance between the second information and the first information; and

updating at least one of the encoder machine-learning network or the decoder machine-learning network using the second measure of distance.

16. The system of claim 15 , wherein the operations further comprise:

calculating a combined distance using the first measure of distance and the second measure of distance; and

updating at least one of the encoder machine-learning network, the decoder machine-learning network, or the channel machine-learning network using the combined distance.

17. The system of claim 15 , wherein the encoder machine-learning network, the decoder machine-learning network, and the channel machine-learning network of the approximated channel are updated concurrently.

18. The system of claim 15 , wherein the encoder machine-learning network or the decoder machine-learning network, and the channel machine-learning network are updated iteratively.

19. One or more non-transitory computer-readable media, storing a computer program, the program comprising instructions that when executed by one or more processing devices cause the one or more processing devices to perform operations comprising:

obtaining first information;

using an encoder machine-learning network to process the first information and generate a first radio-frequency signal;

transmitting the first radio-frequency signal through a first communication channel;

determining a second radio-frequency signal that represents the first radio-frequency signal having been altered by transmission through the first communication channel;

simulating transmission of the first radio-frequency signal over a second communication channel implementing a channel machine-learning network, the second communication channel representing a model of the first communication channel;

determining a simulated radio-frequency signal that represents the first radio-frequency signal having been altered by simulated transmission through the second communication channel;

calculating a first measure of distance between the second radio-frequency signal and the simulated radio-frequency signal; and

updating the channel machine-learning network using the first measure of distance.

20. The media of claim 19 , wherein the operations further comprise:

using a decoder machine-learning network to process the simulated radio-frequency signal and generate second information that is a reconstruction of the first information;

calculating a second measure of distance between the second information and the first information; and

updating at least one of the encoder machine-learning network or the decoder machine-learning network using the second measure of distance.

21. The media of claim 20 , wherein the operations further comprise:

calculating a combined distance using the first measure of distance and the second measure of distance; and

updating at least one of the encoder machine-learning network, the decoder machine-learning network, or the channel machine-learning network using the combined distance.

22. The media of claim 20 , wherein the encoder machine-learning network, the decoder machine-learning network, and the channel machine-learning network of the approximated channel are updated concurrently.

23. The media of claim 20 , wherein the encoder machine-learning network or the decoder machine-learning network, and the channel machine-learning network are updated iteratively.