Recurrent Expansion Framework for Deep Regression with PCA-Guided MLP Augmentation

This repository provides MATLAB implementation of a simple but extensible Recurrent Expansion (RE) framework for supervised regression tasks. The approach iteratively enhances learning by feeding prior model behavior (predictions and internal representations) back into the input. It utilizes a Multi-Layer Perceptron (MLP) with a single hidden layer and incorporates Principal Component Analysis (PCA) to compress intermediate feature representations (φ) for more efficient augmentation. Key components include: basic_recurrent_expansion_mlp.m: Core RE function that trains an MLP over multiple rounds, using PCA-reduced hidden features and prior predictions to augment the input. generate_sinusoidal.m: A helper function to generate a synthetic, noisy sinusoidal regression dataset for demonstration and testing. main.m: An example script that visualizes the dataset, trains the RE model, and plots Mean Squared Error (MSE) across iterations. Default parameters: Hidden layer size = 10 PCA variance threshold = 0.5 (i.e., retain 50% variance) RE rounds = 100 (modifiable) This code can be extended to other architectures and real-world datasets. The included sinusoidal dataset and example results highlight how recurrent expansion enables progressive refinement and self-reflective learning.