Neural Network in Verilog - MNIST Classifier Implementation on DE1-SoC

This project demonstrates a fully functional neural network implemented in Verilog to classify handwritten digits from the MNIST dataset. The design is deployed on an Altera DE1-SoC FPGA board, utilizing PyTorch-trained weights that are quantized and integrated into the hardware.

---

Features

4-Layer Neural Network

• Input: 784 features (28x28 image pixels).
• Architecture: Fully connected layers with ReLU activation.
• Output: 10-class probability distribution (digits 0–9).

State Machine

• Controls layer execution (matrix multiplication, ReLU, and argmax operations).
• Implements an efficient pipeline for sequential processing of neural network layers.

Visualization

• VGA Output: Displays an interactive 28x28 drawing grid.
• Enables real-time testing with predictions displayed on the FPGA.

User Interaction

• Push-Button Controls: Navigate and draw on the grid.
• Seven-Segment Display: Outputs classification results.

---

System Architecture

Neural Network Core

• Sequentially executes:
  • Matrix multiplications for fully connected layers.
  • ReLU activation functions.
  • Argmax operation for final classification.

Memory Management

• Separate memory blocks are used for:
  • Input image data.
  • Weights of each layer.
  • Intermediate results for matrix multiplications and activations.

Drawing Grid

• Users can draw digits on a 28x28 grid using arrow keys.
• The drawn image is processed through the neural network for classification.

---

Training and Deployment Workflow

Training in PyTorch

1. Train the neural network on the MNIST dataset using PyTorch.
2. Quantize weights to 32-bit signed integers for compatibility with Verilog.
3. Export quantized weights to .mif (memory initialization file) format.

FPGA Implementation

• Verilog modules implement core operations:
  • Matrix Multiplication: Handles dot product calculations for each layer.
  • ReLU Activation: Implements element-wise ReLU functionality.
  • Argmax: Determines the class with the highest probability.
• Synthesis and deployment are performed using Intel Quartus Prime.

---

Key Modules

• matrix_multiply: Handles dot product calculations for fully connected layers.
• relu: Implements the ReLU activation function.
• argmax: Identifies the predicted digit with the highest probability.
• VGA Interface: Generates VGA signals for the 28x28 grid visualization.
• State Machine: Controls transitions between neural network layers and interactive features.

---

Visualization

• Sample Outputs:
  • 
  • 

---