This folder contains four pre-trained models that can be loaded to test the Bidirectional Associative Memory (BAM) trained using the following strategies:

SRA: Subspace Rotation Algorithm

SAME: Same-direction alignment of GPA (Gradient-Pattern Alignment) and OWM (Orthogonal Weight Matrix) regularizers

DIFF: Different-direction alignment of GPA and OWM regularizers

ORTH: OWM regularizer only

These models are provided to support the evaluation of robustness and retrieval accuracy, as discussed in our paper.

How to Use
You have two options:

1. Load Pre-trained Models
You can directly load any of the four pre-trained models to evaluate performance without retraining. This is ideal for quickly reproducing the reported results.

2. Train Models Yourself
Alternatively, you can run the training scripts provided to train each BAM variant from scratch. Simply execute the .ipynb file step by step.

Note: Due to the inherent stochastic nature of neural network training (e.g., weight initialization, mini-batch ordering), your results may vary slightly from run to run. However, the overall conclusions of the paper remain valid and reproducible.