Reverse annealing for nonnegative/binary matrix factorization

Area of Science:

• Quantum computing
• Machine learning
• Numerical analysis

Background:

• Quantum annealing is a promising approach for optimization problems.
• Its application in matrix factorization shows potential but faces performance limitations.
• Forward annealing provides quick, approximate solutions but plateaus rapidly.

Purpose of the Study:

• To investigate the efficacy of reverse annealing within quantum annealing for matrix factorization.
• To enhance the performance of nonnegative/binary matrix factorization algorithms.
• To improve solution accuracy and refinement compared to forward annealing.

Main Methods:

• Utilized reverse annealing as a subroutine in quantum annealing.
• Applied the combined forward and reverse annealing approach to matrix factorization problems.
• Compared performance against forward annealing alone across various run times.

Main Results:

• Reverse annealing significantly refines solutions through local searches after initial global search.
• The combination of forward and reverse annealing demonstrated improved performance.
• Performance gains were observed for all run times except the shortest.

Conclusions:

• Reverse annealing enhances quantum annealing for matrix factorization by enabling local search refinement.
• The hybrid forward-reverse annealing strategy offers superior performance for nonnegative/binary matrix factorization.
• This method presents a more effective quantum approach for matrix factorization tasks.