Competitive Decomposition-Based Multiobjective Architecture Search for the Dendritic Neural Model.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Dendritic Neural Models (DNMs) offer computational speed advantages due to logic circuit implementation and binary calculations.
• Optimizing DNM architecture is crucial for enhancing computational speed but presents a large-scale multiobjective optimization problem (LSMOP).
• Conventional multiobjective evolutionary algorithms (MOEAs) struggle with LSMOPs due to irregular Pareto fronts, objective discontinuity, and population degeneration.

Purpose of the Study:

• To develop a novel multiobjective evolutionary algorithm (MOEA) for optimizing dendritic neural model (DNM) architectures.
• To address the limitations of existing MOEAs in handling large-scale multiobjective optimization problems (LSMOPs) specific to DNM architecture search.
• To improve the computational speed and classification performance of DNMs through optimized architecture.

Main Methods:

• Proposed a novel competitive decomposition-based MOEA.
• Decomposed the LSMOP into constrained subproblems with overlapping objective spaces.
• Utilized overlapping regions for environmental selection to propagate selection pressure across the population.

Main Results:

• The proposed algorithm demonstrated superior optimization ability compared to state-of-the-art MOEAs.
• Optimized DNMs achieved competitive classification performances.
• Hardware implementations of the optimized DNMs also showed strong performance.

Conclusions:

• The novel competitive decomposition-based MOEA effectively optimizes DNM architectures for improved speed and accuracy.
• The algorithm overcomes limitations of conventional MOEAs in complex optimization tasks.
• This work advances the practical application of DNMs in machine learning and hardware.