Frankenstein Neural Web

⚡ Pre-loaded Datasets

Quick Start with Built-in Datasets:

Select from three classic machine learning problems to start experimenting immediately:

XOR Problem: Classic non-linear classification problem with 2 binary inputs
Simple Linear Regression: Predict continuous values with a linear relationship
Iris Setosa Classification: Binary classification using the famous Iris dataset

💡 Tip: Pre-loaded datasets are perfect for testing different activation functions and network configurations without preparing your own data.

⚡ Custom Data Preparation

CSV Format Requirements:

First row must contain column headers
1-10 input features (columns), plus 1 output column
Last column is treated as the output/target variable
Supports both numeric and categorical (string) data
No empty cells - all values must be present
Use comma (,) as delimiter

Data Types:

Numeric: Any decimal or integer values (e.g., 1.5, 42, -3.14)
Categorical: Text labels (e.g., "red", "blue", "yes", "no")
Mixed: Combine both types in the same dataset

💡 Tip: The system automatically detects data types and encodes categorical values as numbers for neural network processing.

⚡ Example Datasets

Download these sample datasets to get started:

1. Numeric Only - House Prices

Predict house prices based on size and bedrooms.

size,bedrooms,price
1200,2,250000
1800,3,350000
2400,4,450000
1500,2,280000
2000,3,380000

2. Categorical Only - Fruit Classification

Classify fruits based on color and size.

color,size,fruit
red,small,apple
yellow,medium,banana
orange,medium,orange
red,large,apple
yellow,large,banana

3. Mixed Data - Customer Purchase Prediction

Predict purchases using age (numeric) and membership (categorical).

age,membership,income,purchased
25,bronze,45000,no
35,gold,75000,yes
45,silver,60000,yes
22,bronze,35000,no
50,gold,95000,yes

⚡ Network Configuration

Customize Your Neural Network:

Activation Function: Choose between Sigmoid, ReLU, or Tanh
- Sigmoid: Classic smooth function, good for binary classification (0-1 output)
- ReLU: Fast and efficient, great for deep networks and regression
- Tanh: Similar to sigmoid but outputs range from -1 to 1
Hidden Layer Size: Adjust from 2 to 20 neurons
- More neurons = more capacity to learn complex patterns
- Fewer neurons = faster training, less prone to overfitting
- Default: 6 neurons (good starting point for most problems)

💡 Tip: Try different configurations! ReLU often trains faster, while Sigmoid is better for classification. Increase hidden neurons for complex patterns.

⚡ Training Process

Step-by-Step Guide:

Load Data: Select a pre-loaded dataset or upload your own CSV file
Configure Network: Choose activation function and hidden layer size
Review Encoding: Check the data summary showing detected types and encodings
Start Training: Click "Train Neural Core" to begin
Monitor Progress: Watch the loss graph and terminal output in real-time
Wait for Completion: Training completes after 1000 epochs

Understanding Training Visualizations:

Loss Graph: Real-time plot showing how error decreases over time
Loss Value: Lower is better - should decrease during training
Accuracy: For classification problems, shows percentage of correct predictions
Weight Heatmaps: Visual representation of learned network weights
- Red = positive weights, Blue = negative weights
- Brighter colors = stronger connections

💡 Tip: Good training shows smooth loss decrease. If loss plateaus early, try a different activation function or increase hidden layer size.

⚡ Making Predictions

How to Use the Prediction Interface:

Wait for Training: Prediction interface appears after training completes
Enter Values: Fill in all input fields with appropriate values
Numeric Inputs: Enter numbers directly (e.g., 1500, 3.5)
Categorical Inputs: Select from dropdown or type known values
Click Predict: Get instant results from the trained network

Understanding Results:

Numeric Output: Shows the predicted number
Categorical Output: Shows the decoded label (e.g., "yes", "apple")
Confidence: Values closer to 0 or 1 indicate higher confidence
Unknown Values: Warning appears if you enter values not in training data

💡 Tip: For best results, use input values similar to those in your training data. The network learns patterns from the data you provide.

⚡ Technical Details

Architecture:

3-layer neural network: Input → Hidden (2-20 neurons) → Output
Configurable activation functions: Sigmoid, ReLU, or Tanh
Gradient descent optimization with backpropagation
Learning rate: 0.1
Training: 1000 epochs with real-time visualization

Performance:

Powered by WebAssembly SIMD for fast computation
Runs entirely in your browser - no server required
All data stays on your device
Optimized with x86-64 assembly instructions
Real-time loss graphing and weight visualization

Data Encoding:

Categorical values automatically converted to numbers (label encoding)
Each unique string gets a sequential integer (0, 1, 2, ...)
Encoding is consistent and reversible
Numeric values pass through unchanged

Metrics & Visualization:

Loss tracking: Mean Squared Error (MSE) for all problems
Accuracy: Percentage correct for classification tasks
Interactive loss graph with epoch-by-epoch plotting
Weight heatmaps showing learned patterns in the network

⚡ FRANKENSTEIN NEURAL WEB ⚡

[ HOW TO USE ]

⚡ Pre-loaded Datasets

⚡ Custom Data Preparation

⚡ Example Datasets

1. Numeric Only - House Prices

2. Categorical Only - Fruit Classification

3. Mixed Data - Customer Purchase Prediction

⚡ Network Configuration

⚡ Training Process

⚡ Making Predictions

⚡ Technical Details

[ DATA INJECTION ]

Pre-loaded Datasets

Upload CSV File

[ NEURAL CORE ACTIVATION ]

Network Configuration

Training Loss

Network Weights

[ INFERENCE ENGINE ]