Source-linked AI summary
The quest for a Quantum Neural Network
M. Schuld, I. Sinayskiy, F. Petruccione
TL;DR
QNN research seeks to combine neural computing’s nonlinear, dissipative dynamics with quantum computing’s linear, unitary dynamics. The article systematically establishes requirements and reviews proposals, finding that none fully qualifies as a QNN model while identifying dissipative open-system approaches as a possible way forward.
Problem
QNN research lacks a coherent model that combines neural-network properties with quantum theory while addressing the incompatibility between nonlinear, dissipative and linear, unitary dynamics.
Method
The article introduces neural and quantum computing, establishes three requirements for potential QNNs, and evaluates existing proposals against them.
Results
None of the competing proposals fully satisfies the requirements for a QNN model or fully exploits both quantum and neural-computing advantages.
Takeaways & Limitations
Open Quantum Neural Networks based on dissipative quantum computing are identified as a possible direction for obtaining neural-like dynamics from open quantum systems.
Takeaways & Limitations
Quantum Associative Memory proposals can store approximately 0.138 patterns in a Hopfield network comparison but are not QNN models in the strict sense because they lack a neural-network foundation.
Abstract
from arXiv · showhide
With the overwhelming success in the field of quantum information in the last decades, the "quest" for a Quantum Neural Network (QNN) model began in order to combine quantum computing with the striking properties of neural computing. This article presents a systematic approach to QNN research, which so far consists of a conglomeration of ideas and proposals. It outlines the challenge of combining the nonlinear, dissipative dynamics of neural computing and the linear, unitary dynamics of quantum computing. It establishes requirements for a meaningful QNN and reviews existing literature against these requirements. It is found that none of the proposals for a potential QNN model fully exploits both the advantages of quantum physics and computing in neural networks. An outlook on possible ways forward is given, emphasizing the idea of Open Quantum Neural Networks based on dissipative quantum computing.
1 Introduction
Quantum Neural Networks aim to combine quantum theory with neural-network properties, but the field remains a collection of proposals without a major breakthrough. The article establishes requirements, reviews existing work, and proposes open-system approaches as a possible direction.
- QNNs combine features of quantum theory with the properties of neural networks, whose units are interconnected and feed signals into one another.
- QNNs are not intended to explain brain function because neural dynamics are macroscopic and estimated decoherence times are 10−13 sec and less.
- Their potential comes from exploiting superposition-based quantum computing and parallel-processed neural computing simultaneously.
- QNN research remains an exotic conglomeration of proposals, and a major breakthrough is still outstanding.
- The article establishes requirements for Hopfield-like QNNs with associative memory and reviews existing literature against them.
- The review finds that no proposal satisfies the requirements and outlines Open Quantum Neural Networks using dissipation to obtain neural-like dynamics.
2 Neural computing
Neural computing uses weighted, often binary neuron units whose network dynamics can converge to attractors. Hopfield networks interpret these attractors as stable energy minima that support associative memory.
- Perceptrons: A perceptron transforms incoming binary signals through synaptic weights and applies an activation function to produce a neuron output.
- Neural-network structure: Artificial neural networks encode binary strings in network firing patterns, with neurons connected by synaptic strengths and thresholds.
- Associative memory: Hopfield Neural Networks use a specified connectivity architecture and retrieve stored states closest to an incomplete input by Hamming distance.
- Energy and attractors: Hopfield dynamics converge toward energy minima, making memorised firing patterns stable attractors that retrieve patterns from initial states.
- Learning and capacity: Hebb’s learning rule sets synaptic weights according to correlations across stored patterns, while storage capacity depends on the patterns and their Hamming distances.
- Graded-response networks: Graded-response Hopfield neurons replace binary step functions with sigmoid activations, while retaining equivalent attractor properties in the reported limit.
3 Quantum computing
Quantum computing prepares, evolves, and measures qubits through linear unitary operations, with superposition enabling parallel computation. Its central tension for QNNs is that real environmental interactions introduce dissipation and decoherence, while dissipative computing may engineer attractor dynamics.
- Qubits: Quantum computing uses qubits, two-level quantum systems described in a two-dimensional Hilbert space with basis states |0⟩ and |1⟩.
- Coherent quantum computing: Quantum gates are unitary linear transformations, and qubit superposition enables parallel exploitation of quantum states.
- Dynamics and environment: Quantum physics preserves probability through linear unitary operators, whereas environmental interactions cause dissipation and decoherence.
- Dissipative quantum computing: Dissipative Quantum Computing treats environmental interaction as a means to engineer quantum evolution through open-system dynamics.
- Relevance to QNNs: Dissipative quantum computing is relevant to QNNs because it permits algorithms based on dynamic attractors and steady states, although the described scheme lacks associative memory.
Requirements for a QNN model
A meaningful QNN must combine associative-memory behavior and neural-computing mechanisms with quantum effects while remaining consistent with quantum theory. The central challenge is reconciling nonlinear, dissipative neural dynamics with linear, unitary quantum evolution.
- A QNN replaces a binary McCulloch-Pitts neuron with a quron, a qubit whose two levels represent active and resting firing states.A network of qurons can therefore occupy superpositions of firing patterns.
- The first requirement is associative memory: binary input strings must lead to stable output configurations representing the closest stored pattern under a distance measure.This supports pattern recognition and other central properties of neural information processing.
- The second requirement is a connection to neural computing through mechanisms such as attractor dynamics, synaptic connections, integrate-and-fire behavior, training rules, or neural-network structure.The requirement is intentionally broad to accommodate varied existing and future QNN approaches.
- The third requirement is that evolution exploit superposition, entanglement, or interference while remaining fully consistent with quantum theory.
- Combining these requirements exposes a fundamental tension between nonlinear, dissipative attractor dynamics and linear, unitary quantum theory.In a perceptron, neural activation uses nonlinear step or sigmoid functions, whereas quantum evolution is constrained by quantum dynamics.
- Measurement can imitate a probabilistic step function, but using it at every update destroys quantum effects and reduces the dynamics to a probabilistic classical perceptron.The text therefore calls for a more advanced approach than simple measurement-based updating.
- Existing proposals have not simultaneously captured associative memory and sufficient proximity to neural networks while exploiting quantum mechanics.
4 Review of existing approaches to QNNs
Existing QNN proposals fall into several branches, including measurement-based interpretations, quantum circuits, quantum associative memories, quantum dots, and quantum perceptrons. The review finds that these approaches remain incomplete against the requirements for a meaningful QNN, despite some promising mechanisms and theoretical advantages.
- QNN research has steadily attracted interest, but the field remains a heterogeneous collection of proposals rather than a coherent model.The cited literature includes multiple approaches and the publication count has remained low while interest has grown.
- 4.2 Quantum circuits and 4.4 Quantum Associative Memory (QAM) models: Quantum-circuit proposals encode neural-network mechanisms in quantum gates, while quantum associative memories reproduce associative-memory properties without copying neural-network dynamics or setup.Quantum associative memories theoretically store 2^N patterns in a qubit system of dimension 2^n + 2, exceeding the approximately 0.138 patterns storable in a Hopfield network of N neurons.
- 4.1 Ideas to interpret the step-function as measurement: Measurement-based approaches interpret neural activation or state updating through quantum measurement, sometimes using wave-function collapse to retrieve stored patterns.Peruˇs’s proposal uses a projection operator analogous to Hebb’s rule and repeated measurements to retrieve a nearby memorised state.
- Overall, the reviewed proposals do not provide a fully developed QNN with the required associative-memory dynamics, although measurement-based approaches are described as the most mature response to the dynamics incompatibility.Additional limitations include missing derivations of characteristic Hopfield dynamics, difficulty maintaining global coherence, and ill-defined Hilbert-space constructions.
- 4.3 Interacting quantum dots: Interacting quantum dots can function as an analog quantum neural computer when suitable input-dependent dynamics and parameters engineer the desired input-to-output mapping.This proposal treats the natural evolution of interacting quantum dots as computationally useful under certain conditions.
- 4.5 Quantum Perceptrons: Quantum perceptron proposals focus on constructing quantum versions of neural basic units, but their learning rules face unresolved conflicts between nonunitary learning and probability-preserving unitary evolution.The reviewed learning update is nonunitary and therefore fails to preserve total probability, whereas a unitary rule would not mirror dissipative learning.
5 Discussion
The paper evaluates QNN proposals against requirements addressing neural input-output behavior, neural-network foundations, and consistency with quantum theory. It concludes that no existing proposal yet provides a coherent QNN model, while highlighting open-system and dissipative approaches as a possible direction.
- 5 Discussion: The framework requires QNNs to satisfy neural input-output behavior, a foundation in neural network theory, and consistency with quantum theory.These requirements are used to evaluate existing proposals.
- 5 Discussion: None of the reviewed proposals satisfies all three requirements or fully exploits both neural-computing and quantum-computing advantages.The paper identifies the incompatibility between nonlinear, dissipative neural dynamics and linear, unitary coherent quantum dynamics as the central difficulty.
- 5 Discussion: A promising direction is to use open quantum systems and dissipative quantum computing to obtain dynamical properties similar to neural networks.The proposed Open Quantum Neural Network perspective treats dissipation as a resource for reproducing neural-like dynamics.
- 5 Discussion: Open Quantum Walks may offer environment-coupled dynamics without requiring global coherence between qurons, but their QNN applications remain unstudied.Their output can also be read by measuring an external degree of freedom without destroying coherence.
- 5 Discussion: Dissipation is presented as relevant both for reproducing neural-computing structure and for broader efforts to identify quantum effects in biological processes.The paper distinguishes these QNN motivations from claims that QNNs explain brain function quantum mechanically.