AI-Driven Arithmetic Logic Units

Traditional neural networks struggle to generalize arithmetic operations beyond the numerical range seen during training. This paper investigates the implementation and evaluation of a Neural Arithmetic Logic Unit (NALU).A specialized neural network module introduced in the paper AI-Driven Arithmetic Logic Units, designed to improve numerical reasoning and arithmetic generalization. The NALU architecture extends the Neural Accumulator (NAC) by incorporating both additive and multiplicative computation pathways, controlled by a learnable gating mechanism. This structure enables the model to learn exact arithmetic operations such as addition, subtraction, multiplication, and division, while maintaining the ability to extrapolate beyond the training range. In this manuscript, the NALU model was implemented using TensorFlow and evaluated on synthetic arithmetic tasks.