Source-linked AI summary
Simulating a perceptron on a quantum computer
Maria Schuld, Ilya Sinayskiy, Francesco Petruccione
TL;DR
Quantum machine learning needs a rigorous perceptron analogue for quantum neural networks, particularly because earlier learning proposals can violate unitarity. This paper uses quantum phase estimation in a unitary circuit to reproduce the classical step function, with O(n) resources and quantum-parallel processing as a future learning direction.
Problem
Quantum neural-network research lacks a rigorous perceptron unit that supports classical-style classification and valid quantum learning procedures.
Method
The paper encodes the normalized weighted input in a quantum phase and applies phase estimation within a unitary quantum perceptron circuit.
Results
The circuit reproduces the classical perceptron result with high probability and requires O(n) resources for fixed precision.
Takeaways & Limitations
The model provides a quantum-neural-network building block and could support superposition-based learning with training vectors processed in quantum parallel.
Takeaways & Limitations
An earlier operator-based learning rule does not preserve unitarity, leaving valid quantum learning procedures as an unresolved challenge.
Abstract
from arXiv · showhide
Perceptrons are the basic computational unit of artificial neural networks, as they model the activation mechanism of an output neuron due to incoming signals from its neighbours. As linear classifiers, they play an important role in the foundations of machine learning. In the context of the emerging field of quantum machine learning, several attempts have been made to develop a corresponding unit using quantum information theory. Based on the quantum phase estimation algorithm, this paper introduces a quantum perceptron model imitating the step-activation function of a classical perceptron. This scheme requires resources in $\mathcal{O}(n)$ (where $n$ is the size of the input) and promises efficient applications for more complex structures such as trainable quantum neural networks.
1 Introduction
Perceptrons classify inputs through weighted signals and a step-activation output, but quantum counterparts have struggled with complete implementations and valid learning rules. This paper introduces a unitary quantum circuit that simulates the classical perceptron while supporting quantum-information processing.
- A perceptron combines n binary inputs through weighted connections and activates its output according to a step function.The inputs take values x_k ∈ {−1, 1}, and weights determine whether the output neuron activates.
- Perceptrons are foundational linear classifiers but cannot classify non-linearly separable functions such as XOR.Multi-layer perceptrons later addressed this limitation by combining several perceptron layers.
- Quantum neural-network research seeks a quantum analogue of the perceptron as a building block for larger quantum learning systems.Existing proposals replace neurons with qubits and weights with unitary operators, but some remain incomplete or insufficiently rigorous.
- A proposed quantum learning rule fails to preserve the unitarity condition required for the weight operators.This unresolved issue limits the validity of that proposal’s implementation.
- The paper introduces a unitary quantum circuit that simulates the classical perceptron’s nonlinear input-output function with few additional resources.The model is intended as a classification device and as a building block for quantum neural networks processing information in superposition.
2 The quantum perceptron algorithm
The quantum perceptron encodes a normalized weighted input in a quantum phase and uses phase estimation to reproduce a classical step activation with high probability. Its resources remain O(n) for fixed precision, while precision grows only logarithmically with the square root of neuron count under random inputs and weights.
- Quantum circuit: The circuit writes the normalized net input into a quantum-state phase, then applies inverse quantum Fourier transformation and phase estimation to recover the activation result.The input register encodes neuron states, while an oracle and weight-dependent unitaries prepare the phase.
- Precision: For n = 10, 100, and 1000 neurons, precisions τ ≥4, τ ≥6, and τ ≥8 respectively yield over 85% probability of reproducing the classical perceptron result.The required precision increases because the normalized-input distribution narrows as neuron count grows.
- Precision: The precision scales as τ ∼ log √n under random binary inputs and weights, so increasing neuron count requires only logarithmic growth in resolution.This scaling follows from the assumed standard deviation σ ∼ 1/√n and a resolution on the order of one tenth of σ.
- Complexity: With fixed precision, the quantum perceptron has O(n) complexity comparable to the classical model; using the random-sampling precision relation gives O(n log^2(√n)).The linear scaling comes from the n input-related operations together with the phase-estimation resources.
- Quantum parallelism: A superposed input set is processed in quantum parallel, producing a superposition of outputs that can be measured or further processed by learning algorithms.This capability is presented as a principal advantage over processing input vectors individually.
3 Application in quantum learning algorithms
Quantum learning algorithms can use the quantum perceptron to replace classical perceptron output computation while preserving the training workflow. The model supplies a quantum classification unit, although maintaining unitary operators remains an unresolved issue for some training approaches.
- Quantum perceptron in training: The quantum perceptron can execute the output-computation step of perceptron training, producing a quantum output that can be compared with the target.Training presents an input vector, computes the actual output, compares it with the desired output, and adjusts the weights; the quantum model can perform the computation step.
- Motivation: Feed-forward neural-network training becomes computationally costly because each output depends indirectly on weights across preceding layers.Gradient-descent backpropagation changes weights toward lower error, but the learning phase remains costly for commonly applied network dimensions.
- Motivation: Quantum neural-network research seeks to reduce training time by exploiting quantum effects, including Grover search and quantum-perceptron weight adjustment.These approaches target optimization of neural-network training, but mature quantum learning algorithms remain under development.
- Limitations: An unresolved challenge is ensuring that operators used in some quantum-perceptron training proposals remain unitary or completely positive and trace non-increasing.This constraint is explicitly identified as an open problem for the relevant operator updates.
- Future applications: Superposition-based learning could present the training set as a superposition, process feature vectors in quantum parallel, and make the scheme independent of training-set size.The cited proposal describes this as a future learning direction built on the quantum perceptron.
4 Conclusion
The paper presents a general quantum-computing procedure for simulating a perceptron's step function with classical-comparable efficiency. It positions the model as a missing component for quantum classification and future superposition-based learning schemes.
- 4 Conclusion: The quantum perceptron provides a general procedure for simulating a perceptron’s step function on a quantum computer with classical-equivalent efficiency.The conclusion identifies this as the paper’s central result.
- 4 Conclusion: The model fills a gap in quantum neural-network research by supporting quantum learning methods that require classical-equivalent classification with quantum information.The stated scope concerns quantum learning methods relying on an equivalent classification operation.
- 4 Conclusion: Future superposition-based learning schemes could process superposed training vectors in quantum parallel for quantum machine-learning applications.This is presented as a future outlook rather than a demonstrated result.