New Quantum Optical Neuron From IIT Patna For Scalable AI
Quantum Optical Neurone
Indian Institute of Technology Patna Researchers Discover a Resource-Efficient Quantum Optical Neurone
Vivek Mehta and Utpal Roy of the Indian Institute of Technology Patna constructed a quantum optical model of an artificial neurone, a crucial step towards leveraging quantum technology to accelerate AI. This methodology will significantly reduce computational resources for advanced AI applications. This work presents a photonic circuit architecture that enhances qubit-based neural networks and may lead to scalable and usable quantum neural networks.
Deep neural networks, the foundation of modern AI, demand a lot of computing power for training and deployment. Large language models with billions of parameters are in these networks.
Quantum processing units (QPUs) use quantum mechanical concepts like entanglement and superposition to calculate faster than traditional systems. To enable quantum hardware implementation of deep neural networks, quantum neural network (QNN) methods reduce processing needs.
Brain-Mimicking Artificial Neurone
An artificial neurone mimics biological neurones by computing the inner product of an input vector and a weight vector and applying a non-linear activation function to output. Artificial neurone quantum models use less computer power than their conventional counterparts. This new quantum optical variant handles continuously-valued incoming data using Magnini et al.'s qubit-based paradigm.
Qubit-based Quantum Neurone Challenges and Solutions
Implementing quantum neurones efficiently requires complex quantum circuit fabrication. Qubit-based models store and scale input and weight vectors using quantum wavefunction relative phases. The quantum fidelity implicitly drives the activation function, and the inner product between these states is determined.
Researchers studied two qubit-based quantum circuit synthesis strategies to build the diagonal unitary operator. These methods generate quantum circuits using basic gates like Pauli-Z rotation and two-qubit controlled Pauli-X (Cnot).
Algorithm I: Structures a matrix M using binary and Grey code representations to create a circuit with alternating Cnot gates. Gates act on an ancilla qubit, the target of all Cnot operations. Algorithm II: Bitwise inner products of traditional binary representations yield the components of Algorithm II's matrix M, an unnormalized Hadamard matrix of dimension. This method applies gates to qubits, presumably Cnot gates, based on the binary representation of phase rotations.
Qiskit, a Python-based quantum computation framework, numerically simulated these qubit-based circuits. A thorough analysis of their circuit costs revealed significant disadvantages for wider use:
Circuit Size: Without multi-qubit controlled gates, Algorithm I's circuit size increases with input dimension N, while Algorithm II's decreases. Circuit Depth: Algorithm I, eliminating multi-qubit controlled gates again, has twice the circuit depth of Algorithm II. Circuit width: Both approaches require the same qubits. Interestingly, the qubit-based paradigm sometimes requires measuring all ‘n’ qubits, which increases resource needs and highlights the need for better alternatives.
The Quantum Optical Neurone Promise
Mehta and Roy presented a quantum optical version of the qubit-based quantum neurone to meet this pressing demand by dramatically reducing quantum resource requirements. Photonic technology is ideal for quantum machine learning algorithms since it works at room temperature, uses less energy, and has longer coherence periods.
The quantum optical paradigm stores information in single photon states within spatial quantum modes (qmodes) like qubits do in computing bases. Unitary operations over a reference state are implemented using an integrated programmable quantum optical framework. These operations include:
Multiport Devices (MD): Beam splitters divide or combine light. The tensor product of spatial qmodes and local phase shifters is phase shifters (PS). The team uses optical circuit synthesis to generate beam splitter transmissivity angles to encode real-valued vectors into quantum optical states in the MD. Layers of beam splitters create a pyramidal quantum optical structure.
Validation and Resource Efficiency
Numerical simulations with Strawberry Fields, a Python-based photonic simulation tool, verified the quantum optical model and synthesis algorithm. Simulations of three- and four-dimensional input data matched explicit computations, proving the model's correctness and applicability.
Comparing circuit costs emphasises the quantum optical neuron's advantages:
The quantum optical neuron's circuit size is similar to the qubit-based circuit, expressed as. The optical circuit's depth is much smaller than qubit-based synthesis techniques' circuits. The width of an optical circuit is log, which is always one less than a qubit circuit. By eliminating costly resources like Cnot gates, linear quantum optical circuits make implementation more realistic. The quantum optical model can construct phase-encoded and real-valued quantum neurones, making it adaptable for quantum brain processing.
Their paper “Quantum optical model of an artificial neuron” shows a reduction in resource requirements compared to qubit-based counterparts, suggesting a promising path towards more effective and scalable quantum neural networks for future AI applications.
