NeuroSight

Visual Neural Network Simulator — a real-time desktop application that visualises a deep transformer-based neural network as an interactive brain. Watch forward passes, backpropagation, short-term memory recall, and explainability highlights in real time.

---

Features

• 6-layer neural network visualisation (384→256→128→64→32→12) with live neuron activity
• Forward pass animation — travelling dots show signal flow through the network
• Backpropagation visualisation — orange pulse shows learning in reverse
• 12-class intent recognition — trained on greeting, question, command, complaint, and more
• Short-term memory — associative recall with teal highlight animation
• Explainability mode — gold highlight traces the decision path
• Real-time loss graph — matplotlib live plot during pre-training
• Semantic encoder — sentence-transformers backend for text understanding
• Interactive UI — PyQt6 desktop interface with brain canvas and controls

---

Installation

Requirements

• Python 3.10+
• CUDA-capable GPU recommended (optional, CPU fallback supported)

Quick Install (WSL / Linux)

git clone https://github.com/defenx-tech/NeuroSight.git
cd NeuroSight
bash install.sh

Manual Install

git clone https://github.com/defenx-tech/NeuroSight.git
cd NeuroSight
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install PyQt6 numpy matplotlib sentence-transformers

---

Usage

python main.py

The application window opens with:

1. Brain canvas — 6-layer neural network visualisation with live neuron activity
2. Input field — type text and press Enter to see the forward pass animate
3. Pre-training — runs automatically on startup (300 epochs, ~30 seconds)
4. Loss graph — real-time plot of training loss

Controls

• Type any sentence and watch the signal propagate through the network
• The network classifies into 12 intent categories
• Short-term memory recalls similar past inputs (teal highlight)
• Explainability mode shows the decision path (gold highlight)

---

Architecture

NeuroSight/
├── main.py                 # Application entry point
├── config.py               # Global configuration (architecture, timing, colors)
├── install.sh              # One-shot environment setup
├── requirements.txt        # Python dependencies
├── core/
│   ├── encoder.py          # Text encoder interface
│   ├── memory.py           # Short-term memory ring buffer
│   ├── network.py          # DeepBrainNetwork model
│   ├── projection.py       # Dimensionality projection
│   ├── semantic_encoder.py # sentence-transformers wrapper
│   └── trainer.py          # Training loop
├── ui/
│   └── main_window.py      # PyQt6 main window
├── visualization/
│   └── ...
├── explainability/
│   └── ...
└── ...

---

Example

# Programmatic usage (import core modules)
from core.network import DeepBrainNetwork
from core.semantic_encoder import SemanticEncoder
import torch

# Load the network
model = DeepBrainNetwork()
model.load_state_dict(torch.load("brain_weights.pt"))
model.eval()

# Encode text
enc = SemanticEncoder()
text = "Hello, how are you today?"
embedding = enc.encode(text)

# Run inference
with torch.no_grad():
    logits = model(torch.tensor(embedding).unsqueeze(0))
    probs = torch.softmax(logits, dim=1)
    intent_idx = probs.argmax().item()

print(f"Predicted intent: {config.CLASS_NAMES[intent_idx]}")

---

Tech Stack

• Python 3.10+
• PyTorch — neural network engine
• PyQt6 — desktop GUI framework
• NumPy / Matplotlib — data handling and plotting
• sentence-transformers — semantic text encoding
• CUDA 12.8 — GPU acceleration (RTX 3050+)

---

License

MIT License — see LICENSE.