Quantum-Cognitive Neural Networks: Assessing Confidence and Uncertainty with Human Decision-Making Simulations

TL;DR

QT-NN integrates quantum tunnelling activation with quantum cognition theory to emulate human decision-making in image classification. Through Fashion-MNIST experiments, QT-NN demonstrates human-like uncertainty handling and comparable or superior performance with substantially faster training than classical nets. The study introduces entropy-based uncertainty measures and distributional divergences (KL, JSD) to quantify model confidence and weight dynamics, revealing distinct training behavior compared with classical networks. The results indicate promise for hybrid quantum-classical AI in uncertain, high-stakes domains such as healthcare and autonomous systems.

Abstract

Modern machine learning (ML) systems excel in recognising and classifying images with remarkable accuracy. However, like many computer software systems, they can fail by generating confusing or erroneous outputs or by deferring to human operators to interpret the results and make final decisions. In this paper, we employ the recently proposed quantum-tunnelling neural networks (QT-NNs), inspired by human brain processes, alongside quantum cognition theory, to classify image datasets while emulating human perception and judgment. Our findings suggest that the QT-NN model provides compelling evidence of its potential to replicate human-like decision-making and outperform traditional ML algorithms.

Figures (7)

• Figure S1: Uncertainty in detecting fresh produce items at a supermarket self-checkout equipped with a machine vision system. Left: The system analysed a transparent plastic bag containing truss tomatoes and identified two possible categories: truss tomato and gourmet tomato, leaving the final selection to the customer. Right: In another test with a bag of Amorette mandarins, the system suggested three potential options: Delite mandarin, Amorette mandarin or Navel orange. Similar results were observed with other visually ambiguous items.
• Figure S2: Schematic representation of the QT-NN architecture. The inset illustrates the effect of quantum tunnelling that is employed as an activation function of the network.
• Figure S3: Outputs generated by the QT-NN (red) and the classical neural network model (blue). The insets show the representative testing images for each classification category.
• Figure S4: JSD and SE figures-of-merit for the QT-NN and the classical model for each item category. Note that the classical SE is zero (to machine accuracy) for the 'Ankle Boot' category.
• Figure S5: (a, b) Distributions of weights between the input layer and the hidden layer (denoted as $W_1$ in the main text), plotted as a function of training iterations for the QT-NN model and the classical model (labelled as 'Class.'), respectively. (c, d) Results of the JSD cross-comparison of the initial (labelled as 'Init.') and trained weight distributions $W_1$ and $W_2$. The shaded areas in the JSD plots quantify the divergence, with the numerical value presented above each panel.