AAI_2025_Capstone_Chronicles_Combined

Breast Tumor Classification Using Quantum Neural Networks

14

 Epochs: several values were attempted but settled on 20 for most attempts as that is

when performance seemed to hit a plateau

 Batch size: powers of 2 from 8 (2 3 ) to 128 (2 7 ) were attempted

 1, 2, 5, and 8 (quantum) hidden layers were trialed



Feature selection

 Two approaches normalization for features

 Different quantum feature mappings, which included attempting various cut off values

when mapping between classical data and quantum states 9

Figure 9

Process of quantum feature mapping, with the final value of |0> representing a single data point.

It is important to note that the quantum feature mapping also played a key role in the

data set used. Primarily this included down sampling, including mapping floating point features

to a single (qu)bit. Consequently, feature mapping resulted in colliding values in many cases.

This meant that the same set of down sampled features would map to different targets. These

instances were discarded from the training and testing data, resulting in a significant amount of

9 First normalizing the classical values to a value between 0 and 1. Then for example mapping values less than 0.25 to |0> , values above 0.75 to |1> , and all values between to the equal superposition of |0> and |1> which is (1/2^0.5)|0> + (1/2^0.5)|0> .

267