ABSTRACT Implementing negative weights poses a core bottleneck in photonic computing systems. Conventional methods depend on differential detection or dual‐channel encoding, which incur substantial hardware overhead and restrict system scalability. Here, we demonstrate an intrinsic optical weight encoding scheme that eliminates the need for balanced detection by exploiting controlled interference within a cascaded microring architecture. Unlike conventional dual‐rail, Mach–Zehnder interferometer (MZI), and single microring resonator (MRR) approaches that rely on off‐chip balanced detection to recover weight‐sign information, our method directly encodes signed weights on‐chip through controlled interference between two nonnegative optical channels. The proposed structure achieves continuous and programmable weight representation over the full range of (–1, 1) without balanced photodetection, featuring a tuning efficiency of 2.857 nm/V, a half‐wave voltage of 1.75 V. The fabricated device exhibits a free spectral range of 10.02 nm, extinction ratio of 15 dB, and achieves a correlation coefficient of 96.46 between theoretical and measured weights. Beyond device‐level validation, we demonstrate the compatibility of the proposed scheme with scalable photonic computing by implementing parallel vector–matrix multiplication and convolutional neural network inference. This work establishes a new route toward hardware‐efficient, fully real‐valued optical computing architectures, removing a key barrier in the practical deployment of large‐scale photonic neural networks.
Cui et al. (Mon,) studied this question.