Logisim_MLP

A complete hardware implementation of a multilayer perceptron (MLP) neural network for MNIST handwritten digit recognition, built with Logisim Evolution.

Overview

This project demonstrates a fully digital circuit that classifies 16×16 binarized MNIST images (0–9) in real time. The network comprises:

• Input layer: 256 inputs (16×16 pixels)
• Hidden layer: 128 neurons with ReLU activation
• Output layer: 10 neurons for digit classes 0–9
• Total parameters: ~34,000 weights

All layers and activation functions are realized in Logisim Evolution, with trained weights stored in ROM modules.

Repository Structure

Logisim_MLP/ ├── MLP_circuit.circ # Main Logisim circuit file ├── weights/ # Generated weight and bias files │ ├── layer1_weights.txt │ ├── layer1_bias.txt │ ├── layer2_weights.txt │ └── layer2_bias.txt ├── python/ # Python scripts for weight generation │ └── generate_weights.py └── README.md # This file

Requirements

• Logisim Evolution ≥ 3.8.0
• Java ≥ 8 (for Logisim)
• Python ≥ 3.8 (for weight generation)

Generating Weights

1. Install Python dependencies (if any): pip install numpy

2. Run the weight-generation script: cd python python generate_weights.py

3. Confirm the output files in weights/:

• layer1_weights.txt
• layer1_bias.txt
• layer2_weights.txt
• layer2_bias.txt

Loading the Circuit

1. Open Logisim Evolution.
2. Go to File → Open and select MLP_circuit.circ.
3. Expand the HiddenLayer subcircuit, right-click the weight ROM, choose Load Image, and load:

• weights/layer1_weights.txt
• weights/layer1_bias.txt

4. Repeat for the OutputLayer:

• weights/layer2_weights.txt
• weights/layer2_bias.txt

Simulation Setup

1. In Logisim, enable ticks: Simulate → Ticks Enabled (Ctrl+K).
2. Set clock frequency: Simulate → Tick Frequency → 1 kHz or higher.

Testing & Debugging

1. Start simulation (Ctrl+K).
2. Use the on-screen joystick to draw a digit on the RGB video input.
3. Set the start input to 1 to trigger classification.
4. After ~30–40 s of simulated time, the predicted digit appears on the 7-segment display.

Step-by-Step Debug

• Reduce frequency to 1 Hz for single-step inspection.
• Use the probe tool to monitor:
• InputLayer outputs (256 pixel values)
• HiddenLayer outputs (128 neuron activations)
• OutputLayer scores (10 class logits)

Author

Lloyd Kwak
GitHub: lloydkwak

License

This project is released under the MIT License. Feel free to use and modify for educational purposes.