Research Project | Maturaarbeit

Neural Predictive Calculator

Evaluating the Arithmetic Capabilities of Neural Networks

An in-depth investigation into why neural networks struggle with simple arithmetic expressions. This project builds and evaluates various neural network architectures—including FNNs, RNNs, attention-based RNNs, and Transformers—to assess their performance on arithmetic tasks.

Try Demo View on GitHub

Architectures

GitHub Stars

Open Source

Research

About the Project

Understanding Neural Networks' Arithmetic Limitations

🎯

Research Goal

Investigate why neural networks often struggle with simple arithmetic expressions and evaluate the capabilities of different architectures.

🔬

Methodology

Build and test various neural network models on arithmetic problems, comparing their performance and understanding their limitations.

📊

Scientific Approach

Comprehensive evaluation with statistical analysis, p-values, and detailed findings documented in research papers.

Neural Network Architectures

Exploring Different Approaches to Arithmetic Tasks

📡

Feed-Forward Neural Networks

Best Performance

Basic neural network architecture with multiple layers. FNN2 demonstrated the best performance among all tested models.

✓ FNN1: Initial implementation
✓ FNN2: Optimized version (best model)
✓ Simple and efficient architecture

Key Files: FNN1.ipynb, FNN2.ipynb, NPC-final.py

🔄

Recurrent Neural Networks

Sequential Processing

Networks with loops that allow information persistence, suitable for sequence-based arithmetic operations.

✓ RNN0: Basic implementation
✓ RNN2: Enhanced version
✓ Handles sequential data effectively

Key Files: RNN0.ipynb, RNN2.ipynb

👁️

Attentional RNN

With Attention

LSTM-based RNN with attention mechanism to focus on relevant parts of the input sequence.

✓ LSTM architecture
✓ Attention mechanism
✓ Enhanced context awareness

Key Files: g4gLSTM.ipynb

⚡

Transformer Architecture

State-of-the-art

Modern transformer architecture with self-attention mechanisms for parallel processing.

✓ Multiple iterations (0, 4, 5)
✓ Self-attention mechanisms
✓ Parallel processing capabilities

Key Files: transformer0.ipynb, transformer4.ipynb, transformer5.ipynb

🚀

Pre-trained Transformers

Fine-tuned

Fine-tuned pre-trained models including Gemma (270M & 1B parameters) and Gemini via Vertex AI.

✓ Gemma 270M (Hugging Face)
✓ Gemma 1B (Hugging Face)
✓ Gemini (Vertex AI)

Key Files: gemma_huggingface.ipynb, big_gemma_huggingface.ipynb, gemini_vertex.ipynb

Try It Yourself

Experience the Neural Predictive Calculator

💻

Live Interactive Calculator

Try the FNN2 model right here! Enter an arithmetic expression and see how the neural network computes it.

Enter Expression (e.g., "1 + 2 - 3")

Or Run Locally

Download and run the model on your own machine:

Terminal

# Install dependencies
pip install -r requirements.txt

# Run the demo
python FNN/NPC-final.py

View Demo Script

📦 Installation

All required Python packages are listed in requirements.txt. The models will run on CPU.

🎮 Usage

The demonstration script allows you to input arithmetic expressions and see how the neural network computes the results.

⚙️ GPU Support

If you have an NVIDIA Jetson device, use requirements_GPU.txt for GPU-accelerated inference.

Documentation

Comprehensive Research Materials

📌 Post-Submission Updates: A new branch with minor adjustments, MIT license, and documentation corrections is available at github.com/AntonStantan/matura/tree/post-submission

📊