Symmetrical Bidirectional Optical Convolutional Neural Networks Based on Smart Pixel Light Modulators: A Theoretical Framework

This work introduces a symmetrical smart-pixel-based bidirectional optical convolutional neural network (Sym-SPBOCNN) that unifies forward and backward propagation within a single geometrically symmetric free-space architecture. By pairing two identical SPOCNN modules side by side, the system allows both propagation directions to share the same lens arrangement, pixel geometry, and imaging distances, thereby eliminating the asymmetric optical paths and complex alignments required in earlier SPBOCNN designs. Each smart-pixel light modulator (SPLM) integrates a photodetector, electronic processor, memory, and light emitter, enabling nanosecond-scale weight updates and high-bandwidth optical modulation. Optical analysis confirms that the symmetric configuration preserves high-definition imaging while supporting kernel sizes up to 46×46 under typical lens-array constraints. Performance estimates show that a single 4K-resolution SPLM layer achieves over 4.1×10¹⁴ MAC/s, and a pipelined ten-layer configuration exceeds 4.1×10¹⁵ MAC/s without additional overhead. The architecture also inherits the reciprocity of SPBOCNNs, enabling efficient optical inference while substantially reducing the complexity of optical design, fabrication, and alignment. These characteristics establish the Sym-SPBOCNN as a compact and scalable bidirectional optical processor suitable for future hybrid electro-optical AI hardware.