Approaches to Avoid Overfitting in a Quantum Perceptron

In this paper, we analyze an existing Quantum Perceptron in terms of its generalization performance and introduce alternative strategies to avoid classification errors, the Circuit Threshold Operator (CTO) and Neuron Threshold Variation (NTV). The adjust of the CTO and NTV parameters increases the probability of the neuron to tolerate differences between input and stored weights allowing some noise level to be acceptable. Experiments were conducted in IBM quantum simulator, showing that our proposals are effective when compared to the original fixed parameter. A hybrid classical-quantum training was executed, using Genetic Algorithm to train the parameters of the model classically and running our model in a quantum simulator. Our solution achieves 78.57% and 76.43% of accuracy on a noisy dataset, while the original neuron approach does not exceed 53.57%.