Quantum Annealing Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep learning neural networks on conventional computing platforms is slow and requires heavy computational resources, making it inefficient for complex tasks like handwriting recognition and image recognition.
Innovation Solution
Configuring a Quantum Annealing (QA) device to act as a quantum neural network by mapping nodes and connections to qubits and couplers, and using a hybrid classical/quantum computing architecture to speed up the training process through quantum-assisted training methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional computing platforms are used to train deep learning neural networks, then the training can be performed with existing infrastructure, but the computational time is extremely long and heavy computational resources are required
Solution Approach 1:
The patent replaces conventional classical computing systems with a quantum annealing system to perform neural network training. The quantum system uses quantum mechanical effects (quantum tunneling and superposition) to explore the energy landscape of the neural network weights, substituting the classical iterative optimization process. This is achieved by mapping the neural network training problem onto a quantum Hamiltonian and using quantum annealing to find the ground state, which corresponds to the optimal weights.
Solution Approach 2:
The patent changes the fundamental parameters of the computing system by transitioning from classical bits to quantum bits (qubits). This parameter change enables the system to exploit quantum mechanical properties such as superposition and entanglement, allowing simultaneous exploration of multiple weight configurations. The quantum annealing process uses temperature-like parameters and quantum coupling strengths to control the optimization trajectory, fundamentally differing from classical optimization algorithms.
2Productivity
If conventional computing platforms are used to train deep learning neural networks, then existing hardware can be utilized, but heavy computational resources and long training durations are required
Solution Approach 1:
The patent substitutes the classical computing architecture with a quantum annealing processor that specialized in optimization problems. The quantum system's inherent parallelism through quantum superposition allows it to evaluate multiple weight configurations simultaneously, dramatically improving training efficiency. The quantum annealer's physical architecture directly implements the energy minimization process, eliminating the need for sequential classical computation iterations.
Solution Approach 2:
The patent creates a quantum copy of the neural network's weight space by mapping each weight parameter to a corresponding quantum coupling between qubits. This quantum representation allows the system to explore the entire weight landscape through quantum fluctuations and tunneling, effectively copying and evaluating numerous potential solutions in parallel without requiring proportional classical computational resources.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Significantly reduces the computational time required for training deep learning neural networks by leveraging quantum properties to generate quantum samples and update weights and biases efficiently, thereby accelerating the training process.
Implementation Method 1
a quantum annealing process is performed on the quantum computer to generate a plurality of quantum samples
Implementation Method 2
where a quotient of an energy functional of the RBM being divided by a scale factor β eff is used as a final Hamiltonian
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
Aspects of the disclosure provide a method for configuring a Quantum Annealing (QA) device. Then QA device has a plurality of qubits and a plurality of couplers at overlapping intersections of the qubits. The method includes mapping a node of a neural network that have a plurality of nodes and connections between the nodes to a qubit in the QA device, and mapping a connection of the neural network to a coupler at an intersection in the QA device where two qubits corresponding to two nodes connected by the connection intersect. The method further includes mapping a node of the neural network to a chain of qubits. In an embodiment, a coupling between qubits in the chain is configured to be a ferromagnetic coupling in order to map the node of the neural network to the chain of qubits.