External Quantum Self-Attention Model

Self-attention mechanisms have revolutionized machine learning domains. Quantum self-attention models have emerged as promising extensions, yet they suffer from quadratic computational complexity like their classical counterparts. To address this, we propose the External Quantum Self-Attention Model (EQSAM), which integrates external memory modules into quantum self-attention, reducing complexity from quadratic to linear. In EQSAM, two sets of fully trainable external quantum modules generate key and value memory states. The model computes similarities only between input queries and these external states, rather than pairwise among all inputs. This lowers computational demands and boosts scalability in quantum machine learning. Experiments on MNIST and Fashion MNIST classification tasks demonstrate that EQSAM achieves comparable or superior performance to pairwise quantum self-attention models with less computation. The number of external modules serves as a key hyperparameter in EQSAM, with an optimal value balancing representational capacity and generalization. Performance approaches saturation near this optimum.