Neural Networks

Machine learning models

Neural Networks

Neural networks are machine learning models that can learn complex patterns in your data for classification and regression tasks.

Types of Tasks

Classification

Predict categories or classes.

Example: Spam vs. not spam, flower species identification.

Regression

Predict continuous values.

Example: House prices, stock prices, temperature forecasting.

Data Requirements

Input Data

  • Numeric features: All input variables must be numeric
  • Clean data: Remove or handle missing values
  • Sufficient samples: More data generally leads to better performance

Target Variable

  • Classification: Categorical variable (2+ classes)
  • Regression: Continuous numeric variable

How to Run

Step 1: Select Neural Network

  1. Choose “Neural Network” from the test dropdown
  2. Select your CSV file

Step 2: Configure the Model

  1. Task Type: Choose Classification or Regression
  2. Dependent Variable: Select your target variable
  3. Independent Variables: Select one or more feature variables
  4. Hyperparameters (optional):
    • Hidden Layer Size: Number of neurons (default: 10)
    • Learning Rate: Step size for training (default: 0.01)
    • Epochs: Training iterations (default: 1000)

Step 3: Run Analysis

Click “Run Analysis” to train your neural network.

Interpreting Results

Model Information

  • Architecture: Input → Hidden → Output layer sizes
  • Task Type: Classification or Regression
  • Hyperparameters: Training configuration used

Performance Metrics

For Classification:

  • Accuracy: Percentage of correct predictions
  • Number of Classes: Total categories in your data

For Regression:

  • MSE (Mean Squared Error): Average squared prediction error
  • RMSE (Root Mean Squared Error): Square root of MSE
  • MAE (Mean Absolute Error): Average absolute prediction error
  • R²: Proportion of variance explained by the model

Training Progress

  • Training Loss: Error on training data
  • Validation Loss: Error on validation data
  • Convergence: Whether the model stopped improving

Model Download

Download Options

  1. Model File (.pkl): Complete trained neural network
  2. Model Info (.json): Architecture and performance details
  3. Python Script: Ready-to-use code for deployment

Using Downloaded Models

import numpy as np
import pickle
import base64

# Load your model
model = SimpleNeuralNetwork(input_size, hidden_size, output_size)
model.load_model(model_base64)

# Preprocess new data
normalized_data = preprocess_data(new_data, X_mean, X_std)

# Make predictions
predictions = model.predict(normalized_data)

Best Practices

Data Preparation

  • Clean your data: Remove missing values and outliers
  • Feature selection: Choose relevant variables
  • Data splitting: Ensure sufficient training data

Model Configuration

  • Start simple: Use default hyperparameters first
  • Experiment: Try different hidden layer sizes
  • Monitor training: Watch for overfitting

Limitations

Current Implementation

  • Single hidden layer: Limited to one hidden layer
  • Fixed architecture: Cannot add custom layers
  • Basic optimization: Uses simple gradient descent

General Limitations

  • Black-box nature: Difficult to interpret decisions
  • Data requirements: Need sufficient training data
  • Computational cost: Training can be time-consuming
  • Overfitting risk: Can memorize training data

Export Options

  • LaTeX: Publication-ready tables
  • CSV: Structured data format
  • Model files: Complete models for deployment

Need help? Check the FAQ or contact support.