SPARC increases the accuracy of photonic neural network accelerators by computing representative vector inner products digitally and offloading smaller residual products to analog hardware. This hybrid approach mitigates errors caused by low-bit quantization and analog noise.
HOW THIS AFFECTS YOU
●
builderThis approach allows you to use power-efficient photonic hardware without the usual hit to inference precision.
●
researcherYou can leverage activation vector similarity to reduce the impact of quantization errors in analog computing.