High-speed feature extraction with integrated microwave neurons

Intelligence at high data rates depends on extracting meaning the instant information arrives. However, while real visual, audio and radio signals are fundamentally analog, most computers still fully digitize them before extracting features, which constrains processing speed to the pace of a digital clock and adds latency and power overhead from analog-to-digital conversion. We find that, instead, an ensemble of coupled waveguides could extract spectral features of incoming signals instantaneously through interactions between frequency modes across tens of gigahertz. This milliwatt-scale Microwave Neural Network (MNN) uses the incoming signals themselves to reconfigure coupling between waveguides, reshaping its spectrum in real time. By instantly expressing each token’s features across many frequencies, the MNN makes the relationships between successive tokens easier to detect, enabling faster decision-making — without digital preprocessing. Also, when the waveguides are driven directly by gigabit-per-second bitstreams, such as 8-bit pixel data, the MNN produces probabilistic-bits whose bias reflects the input pattern. This acts like an analog form of dithering with lower overhead, where controlled randomness preserves fine detail using fewer bits. As a result, the readout can sample at only one-eigth of the bit rate while still retaining the distinguishing features of the original pattern. This could let satellites downlink richer imagery within tight bandwidth limits and enable radar sensors to classify threats from radio-frequency signals as they arrive—embedding context directly into clockless front-end circuitry.