Learning Representations by Back-Propagating Errors (1986)

An independent PyTorch implementation of the classic multi-layer feedforward neural network and custom batch gradient descent algorithm proposed by Rumelhart, Hinton, and Williams (1986). This project replicates the Symmetry Detection task using the exact architectural constraints and learning parameters outlined in the original paper.

---

Repository Structure

your-repo/
├── PAPER_NOTES.md    # Summary of the paper's claims, method, and metrics
├── README.md         # Setup and execution instructions
├── task3.py          # Core model architecture, custom loss, and optimizer
└── run_experiment.py # Dataset generation and the full-batch training loop

System Requirements & Dependencies
This project requires a standard Python 3 environment and the torch library.

Python: 3.8+

PyTorch: 2.0+

Installation
To install the required dependencies, run the following command in your terminal:
bash: 
pip install torch

How to Run the Code
The dataset generation and training process are entirely automated within a single execution script. To start the training loop, run:
bash:
python run_experiment.py

What to Expect During ExecutionThe script programmatically builds all 64 unique 6-bit binary sequences and maps them to their respective symmetry targets.It initializes a network containing exactly 2 hidden units, tracking learning across 1,500 full-batch sweeps using the paper's explicit parameters ($\epsilon = 0.1$, $\alpha = 0.9$).Progress updates, tracking error convergence and classification accuracy, will print out to the console every 100 sweeps.Verification & Expected ResultsThe implementation has been successfully verified on the 6-bit Symmetry Detection problem:Target Dataset: 64 total possible bit patterns (Full-Batch learning).Observed Metrics: Consistent convergence achieving an overall performance accuracy of 87.50% (perfectly classifying 56 out of 64 patterns) within 1,500 sweeps.