Neural Network Visualizer – MNIST Classification

!!! You need to map mnist_test.csv (feature values 0-255) to mnist_mapped.csv (0-1) !!!

Go to mnist_test.ipynb and run the first few cells. then run the main (NEURAL2-TORCH.py)

This project is an interactive visualizer for a multi-layer perceptron (MLP) that classifies handwritten digits from the MNIST dataset.
It trains the model in real-time while showing:

• The weights between neurons, updating as learning progresses
• The pixel input grid for the current sample
• The model’s predictions for each digit class
• The target label for the sample being trained

---

Features

• Live Training Visualization – Watch weights change after every batch.
• Pixel Grid Display – All 784 input pixels shown in sync with the current sample.
• Neural Network Diagram – Neurons and connections drawn in layers.
• Prediction Output – Real-time output activations and predicted digit.
• Interactive Controls – Start/stop training on key press.

---

Model Architecture

The model is a simple MLP (Multi-Layer Perceptron) with ReLU activations:

Input Layer: 784 neurons (28x28 pixels) Hidden Layer 1: 128 neurons, ReLU Hidden Layer 2: 64 neurons, ReLU Output Layer: 10 neurons (digits 0–9), Softmax

---

Requirements

• Python 3.8+
• PyTorch
• Pygame
• Matplotlib
• NumPy

Install dependencies:

pip install torch pygame matplotlib numpy pandas scikit-learn

Controls

T – Toggle training on/off

ESC / Close window – Exit

How It Works

Loads the MNIST dataset using PyTorch’s DataLoader.

Builds an MLP for classification.

Updates weights in real-time and visualizes them as colored lines.

Displays current pixel input in a 28x28 grid.

Shows prediction scores and target label for each sample.

License

This project is open-source under the MIT License.