Lossy compression of matrices by black box optimisation of mixed integer nonlinear programming.

Area of Science:

• Edge Computing
• Machine Learning
• Optimization Algorithms

Background:

• Data compression is crucial for machine learning models in resource-constrained edge computing environments, especially for complex tasks like autonomous driving.
• Existing lossy compression methods decompose matrices but face optimization challenges due to simultaneous real and integer variable optimization.

Purpose of the Study:

• To improve the optimization of lossy compression for matrix data in edge computing.
• To address the simultaneous optimization of integer and real variables in matrix decomposition.

Main Methods:

• Utilized recently developed black-box optimization (BBO) algorithms.
• Integrated an Ising solver for binary variables to enhance optimization.
• Applied the algorithm to solve mixed-integer programming problems (linear in real, non-linear in integer variables).

Main Results:

• Successfully improved the optimization of matrix data compression for edge computing applications.
• Demonstrated the applicability of BBO with Ising solvers to mixed-integer programming problems.
• Provided a framework for further development of BBO techniques.

Conclusions:

• The proposed BBO with Ising solver approach offers an effective solution for data compression in edge AI.
• This method enhances computational efficiency for complex machine learning tasks.
• Further research can explore variations in Ising solvers and BBO strategies for optimal performance.