Radio-frequency (RF) sensing underpins applications ranging from radar and wireless communication to biomedical and quantum measurement, where detection sensitivity at low signal-to-noise ratio (SNR) directly limits the achievable range, resolution, and information capacity. Machine learning has been widely applied to enhance sensing performance, but predominantly as post-detection analysis of already-acquired data. At low SNR, however, the information missing from acquired data cannot be recovered downstream, making pre-detection processing essential. Here we introduce microring perceptron (MiRP) sensing, a framework that exploits RF-photonic transduction in a microring resonator, where a programmable optical pump performs three-wave mixing to implement a learned mapping from the incoming RF signal to optical feature signals prior to detection. The transducer front end and digital neural-network back end are jointly optimized within a single end-to-end training pipeline, allowing the sensing system to learn an encoding that preserves task-relevant information through detection. Across benchmark tasks, MiRP sensing achieves substantially higher task performance than conventional processing at low input RF power levels where detection noise dominates. The algorithmic gain reported here offers an orthogonal axis of improvement that composes with, and amplifies the impact of, existing and future advances along complementary dimensions, including hardware efficiency and quantum-enhanced optical sensing.