On the multi-objective hyperparameter optimization of the weighted entropic associative memory

This paper presents an application of automatic algorithm configuration to optimize the performance of the Weighted Entropic Associative Memory (W-EAM). This model holds declarative but distributed representations of remembered objects. We apply two state-of-the-art hyperparameter tuning methods-SMAC and SMS-EMOA-to improve W-EAM's recognition accuracy across three domains of increasing complexity: digit recognition, character recognition, and phone recognition in Mexican Spanish. Each domain presents unique challenges in memory structure, data representation, and learning dynamics. We evaluate a multi-objective optimization setting, focusing on the trade-off between precision and recall. Our experiments show that optimized configurations consistently outperform the baseline model, with the largest gains observed in the phone domain. Additionally, a cyclical optimization-learning scheme further enhances performance by iteratively improving both model parameters and training data quality. These results highlight the potential of automated configuration methods in advancing adaptive memory systems and providing practitioners with a range of parameters to use on W-EAM.