Transforming Traditional Neural Networks into Neuromorphic Quantum-Cognitive Models: A Tutorial with Applications

TL;DR

This work addresses the barrier to quantum-inspired AI by showing how traditional neural networks can be transformed into QT-based neuromorphic models that run on a laptop. The authors replace conventional activations with a quantum-tunnelling mechanism while preserving Softmax, enabling faster training and emergent cognitive-like properties across FNNs, RNNs, ESN, and BNNs. They demonstrate strong results on MNIST, Fashion-MNIST, sentiment analysis, and Mackey-Glass forecasting, with notable speedups and robust uncertainty handling. The study also discusses cognitive human-machine teaming and multimodal extensions, highlighting practical implications for defence, healthcare, and adaptive AI, and points toward scalable, hardware-integrated future work.

Abstract

Quantum technologies are increasingly pervasive, underpinning the operation of numerous electronic, optical and medical devices. Today, we are also witnessing rapid advancements in quantum computing and communication. However, access to quantum technologies in computation remains largely limited to professionals in research organisations and high-tech industries. This paper demonstrates how traditional neural networks can be transformed into neuromorphic quantum models, enabling anyone with a basic understanding of undergraduate-level machine learning to create quantum-inspired models that mimic the functioning of the human brain -- all using a standard laptop. We present several examples of these quantum machine learning transformations and explore their potential applications, aiming to make quantum technology more accessible and practical for broader use. The examples discussed in this paper include quantum-inspired analogues of feedforward neural networks, recurrent neural networks, Echo State Network reservoir computing and Bayesian neural networks, demonstrating that a quantum approach can both optimise the training process and equip the models with certain human-like cognitive characteristics.

Figures (9)

• Figure S1: (a) Schrödinger’s cat thought experiment. A cat is placed in a sealed, opaque box containing a radioactive atom, a Geiger counter, a vial of poison and a hammer. If the atom decays, the Geiger counter triggers the hammer to release the poison, killing the cat. Until the box is opened and observed, quantum mechanics suggests the cat exists in a superposition of being both alive and dead. (b) Illustration of a projective measurement of a qubit $|\psi\rangle$ using the Bloch sphere, where measurement collapses the qubit from a superposition to a definite state. (c) The double-slit experiment, which demonstrates quantum interference, showing that particles such as electrons behave as waves, creating an interference pattern until observed. (d) Illustration of wavefunction collapse triggered by detection using an electron detector.
• Figure S2: (a.i)--(a.iv) Instantaneous snapshots of an energy wave packet modelling the interaction of an electron with a double-slit structure. (b.i)--(b.iv) Instantaneous snapshots of an energy wave packet modelling the tunnelling of an electron through a continuous potential barrier. The false-colour scale of the images encodes the computed probability density values.
• Figure S3: Schematic representation of a generic neural network model, illustrating the replacement of the traditional ReLU activation function with the physical QT effect. Other types of activation functions can be substituted in a similar manner. In this paper, the Softmax activation function remains unaltered. The proposed replacement approach has been demonstrated to be effective across all neural network models examined in this study.
• Figure S4: Fourier spectra of the outputs of (a, e) ReLU, (b, f) Sigmoid, (c, g) Identity and (d, h) QT functions activated by a sinusoidal wave signal at a frequency of 16 Hz. Note the highest number of nonlinearly generated higher-order harmonic in the spectrum of the QT function.
• Figure S5: Example of classifications from a randomly selected subset of the MNIST testing dataset produced by the QT-feedforward neural network. The labels above each panel indicate the predicted categories alongside the ground truth labels.