Co-frequency signal separation at 235K parameters: complex attention cuts collapse and keeps phase

A complex-valued convolutional network with only 235,000 parameters can separate two co-frequency communication signals and falls into output collapse less often during training. Earlier communication-signal separation networks either treat I/Q as two independent real channels, discarding the phase coupling that carries modulation structure, or stack attention modules onto a complex backbone without any mechanism for real-imaginary correlation. On synthetic 5 Hz co-frequency mixtures covering all 16 modulation pairs of BPSK, QPSK, 8PSK and 16QAM at 10 dB SNR, the model reaches a five-seed mean SDR of 2.31 dB with 0.62 dB standard deviation; the parameter-matched no-attention complex CNN reaches 1.81 dB with 0.48 dB standard deviation and collapses on four of five seeds, against two of five for this model.

The gain comes from a channel-attention block built for complex feature maps: it pools real and imaginary statistics, then emits one real weight that scales each channel. The block, complex squeeze-and-excitation (C-SE), applies global average pooling over time separately to the real and imaginary parts, concatenates the descriptors, passes them through a shared two-layer perceptron with sigmoid gating, and averages the two halves of the 2C output into a single real channel weight. The network is one complex convolutional encoder, four complex residual blocks carrying C-SE, and one complex convolutional decoder that produces two complex masks multiplied element-wise with the observed mixture; training uses a loss split evenly between mean squared error and SI-SDR on a single 8 GB RTX 4060.

The demonstrated setting is synthetic two-source co-frequency mixtures: 16 kHz sampling, 4096-sample signals, root-raised-cosine shaping, three-tap multipath, per-source carrier offsets within plus or minus 5 Hz, and symbol-aligned pulse trains, meaning coarse symbol timing has already been acquired. The authors list relaxing symbol alignment, extending beyond two sources, dilated convolutions for longer temporal modelling, and validation on over-the-air and public I/Q datasets as future work.

The authors explicitly make no claim of superiority or equivalence: paired SDR differences span minus 0.48 dB to plus 0.50 dB, none reaches alpha = 0.05, and with five pairs the exact test has a structural floor of 0.0625. Output collapse appears per seed in most evaluated models, and C-SE only lowers the rate from four of five seeds to two of five. In the scale-mode ablation the complex-weight variant confines phase to the first quadrant, which the authors note is not a fair test of whether phase matters.