Autoencoder-based error correction coding for low-resolution communication
View Patent ↗Various embodiments of the present technology provide a novel deep learning-based error correction coding scheme for AWGN channels under the constraint of moderate to low bit quantization (e.g., one-bit quantization) in the receiver. Some embodiments of the error correction code minimize the probability of bit error can be obtained by perfectly training a special autoencoder, in which “perfectly” refers to finding the global minima of its cost function. However, perfect training is not possible in most cases. To approach the performance of a perfectly trained autoencoder with a suboptimum training, some embodiments utilize turbo codes as an implicit regularization, i.e., using a concatenation of a turbo code and an autoencoder.
1. A transmitter in a communication system comprising:
an encoder to receive a stream of information bits at a rate and produce an encoded bit stream as output that adds redundancy to the stream of information bits;
a digital modulator communicably coupled to the encoder and configured to receive the encoded bit stream from the encoder and produce complex data symbols as output;
a neural network communicably coupled to the digital modulator and configured to receive the complex data symbols and produce an output vector using a vector of one or more of the complex data symbols, the output vector having a dimension greater than or equal to a dimension of the vector of one or more of the data symbols;
a serial to parallel converter communicably coupled to the digital modulator and the neural network and configured to receive the complex data symbols and produce the vector of one or more of the complex data symbols, the vector of one or more of the complex data symbols is fed as input to the neural network; and
one or more digital to analog convertors communicably coupled to the neural network and configured to receive the output vector, and produce an analog waveform to be transmitted over a communication channel.
2. The transmitter of claim 1 , wherein the neural network includes one or more layers each associated with weights.
3. The transmitter of claim 1 , wherein the digital-to-analog converter has a quantization resolution less than or equal to six bits.
4. The transmitter of claim 1 , wherein the encoder is a turbo encoder, polar encoder, or a low-resolution parity-check (LDPC) encoder.
5. The transmitter of claim 1 , wherein the digital modulator includes a quadrature amplitude modulation (QAM) modulator or phase-shift keying (PSK) modulator.
6. The transmitter of claim 1 , wherein the neural network includes a single layer neural network or a deep neural network.
7. A method comprising:
encoding a stream of information bits to produce a corresponding encoded bit stream; modulating the corresponding encoded bit stream to generate a first sequence of complex data symbols;
converting the first sequence of complex data symbols into a plurality of first vectors of complex data symbols;
generating, using the first sequence of complex data symbols, a second sequence of complex data symbols having a dimension greater than or equal to a dimension of the first sequence of complex data symbols, the second sequence of complex data symbols different from the first sequence of complex data symbols;
generating, for each first vector, a corresponding second vector of the second sequence of complex data symbols having dimension greater than or equal to a dimension of the first vector; and
converting the second sequence of complex data symbols into one or more analog signals for transmission over a communication channel.
8. The method of claim 7 , wherein generating the second sequence of complex data symbols includes using a neural network having one or more layers.