Neural Network Function Approximation Experiments

Overview

This repository contains PyTorch implementations for neural network-based function approximation experiments. Each folder corresponds to a specific experiment setup as presented in the accompanying research paper.

Environment Requirements

• Python 3.8+
• Core Dependencies:
  • Torch: 2.6.0+cu124
  • Scipy: 1.15.2
  • Numpy: 2.2.3

Project Structure

experiments/

• linearApprox: Approximation to three-dimensional linear functions;
• pieceApprox: Approximation to piece-wise functions;
• sinApprox: Approximation to sin(2x) and sin(2x) + sin(6x) + sin(10x) functions;
• linear_pretrain: Pre-training on one-dimensional linear function approximation
• sinx_interpolation: Sine interpolation experiments;
• func_img: Image processing experiments;
• func_video: Video processing experiments.

Experiment Folder Structure

Each experiment folder contains:

• model_x1/, model_x2/ - Trained model parameters
• results/ - Experimental results and loss values
• main.py - Main execution script
• data_generation.py - Training data generation
• NNnetwork.py - Neural network architecture
• train_class.py - Training procedure implementation
• save_data.py - Data persistence utilities
• save_test_loss.py - Loss recording functionality

Quick Start

1. Install dependencies:

pip install torch==2.6.0 scipy==1.15.2 numpy==2.2.3