Hello, I'm

RAVIN GUPTA

Neural Network Demonstration

Interactive Deep Learning Model with 5 Layers: 3 Input → 4 Hidden → 3 Hidden → 2 Hidden → 2 Output Nodes

Network Architecture

A simple feedforward neural network with adjustable parameters

Input Layer

0.50
0.30
0.80

Hidden Layer 1

0.00
0.00
0.00
0.00

Hidden Layer 2

0.00
0.00
0.00

Hidden Layer 3

0.00
0.00

Output Layer

0.00
0.00

Input Values

0.50
0.30
0.80

Target Values

0.80
0.20
0.10
Epoch: 0
Loss: 0.0000

Mathematical Formulas & Calculations

Real-time view of all mathematical operations

Activation Function (Sigmoid)

Formula: σ(x) = 1 / (1 + e^(-x))

This function maps any real number to a value between 0 and 1

Current Bias Values

Hidden Layer 1:

H1-1: 0.100

H1-2: -0.200

H1-3: 0.300

H1-4: 0.000

Hidden Layer 2:

H2-1: 0.200

H2-2: -0.100

H2-3: 0.400

Hidden Layer 3:

H3-1: 0.100

H3-2: -0.300

Output Layer:

Out-1: 0.200

Out-2: -0.100

Hidden Layer 1 Calculations

H1-1:

sum = 0.50 × 0.40 + 0.30 × 0.20 + 0.80 × 0.60 + bias

sum = 0.7400 + 0.100 = 0.8400

σ(0.8400) = 0.0000

H1-2:

sum = 0.50 × 0.70 + 0.30 × -0.30 + 0.80 × 0.10 + bias

sum = 0.3400 + -0.200 = 0.1400

σ(0.1400) = 0.0000

H1-3:

sum = 0.50 × -0.20 + 0.30 × 0.80 + 0.80 × -0.40 + bias

sum = -0.1800 + 0.300 = 0.1200

σ(0.1200) = 0.0000

H1-4:

sum = 0.50 × 0.50 + 0.30 × 0.10 + 0.80 × 0.90 + bias

sum = 1.0000 + 0.000 = 1.0000

σ(1.0000) = 0.0000

Hidden Layer 2 Calculations

H2-1:

sum = 0.00 × 0.30 + 0.00 × 0.80 + 0.00 × -0.10 + 0.00 × 0.50 + bias

sum = 0.0000 + 0.200 = 0.2000

σ(0.2000) = 0.0000

H2-2:

sum = 0.00 × -0.50 + 0.00 × 0.20 + 0.00 × 0.90 + 0.00 × -0.30 + bias

sum = 0.0000 + -0.100 = -0.1000

σ(-0.1000) = 0.0000

H2-3:

sum = 0.00 × 0.70 + 0.00 × -0.60 + 0.00 × 0.40 + 0.00 × 0.10 + bias

sum = 0.0000 + 0.400 = 0.4000

σ(0.4000) = 0.0000

Hidden Layer 3 Calculations

H3-1:

sum = 0.00 × 0.40 + 0.00 × 0.70 + 0.00 × -0.20 + bias

sum = 0.0000 + 0.100 = 0.1000

σ(0.1000) = 0.0000

H3-2:

sum = 0.00 × -0.60 + 0.00 × 0.30 + 0.00 × 0.90 + bias

sum = 0.0000 + -0.300 = -0.3000

σ(-0.3000) = 0.0000

Output Layer Calculations

Output 1:

sum = 0.00 × 0.60 + 0.00 × 0.20 + bias

sum = 0.0000 + 0.200 = 0.2000

σ(0.2000) = 0.0000

Output 2:

sum = 0.00 × -0.40 + 0.00 × 0.80 + bias

sum = 0.0000 + -0.100 = -0.1000

σ(-0.1000) = 0.0000

Loss Calculation (Mean Squared Error)

Formula: Loss = 0.5 × Σ(predicted - target)²

Loss = 0.5 × [(0.0000 - 0.80)² + (0.0000 - 0.20)²]

Loss = 0.5 × [(-0.8000)² + (-0.2000)²]

Loss = 0.5 × [0.640000 + 0.040000]

Loss = 0.000000

Backpropagation: Loss Gradients for Each Layer

Output Layer Gradients

Output 1:

Error = predicted - target = 0.0000 - 0.80 = -0.8000

Sigmoid'(x) = output × (1 - output) = 0.0000 × 1.0000 = 0.0000

Gradient = -0.8000 × 0.0000 = 0.000000

Output 2:

Error = predicted - target = 0.0000 - 0.20 = -0.2000

Sigmoid'(x) = output × (1 - output) = 0.0000 × 1.0000 = 0.0000

Gradient = -0.2000 × 0.0000 = 0.000000

Hidden Layer 3 Gradients

H3-1:

Weighted error sum = (-0.800 × 0.000 × 0.60) + (-0.200 × 0.000 × -0.40)

Weighted error sum = 0.000000

Sigmoid'(x) = 0.0000 × 1.0000 = 0.0000

Gradient = 0.000000 × 0.0000 = 0.000000

H3-2:

Weighted error sum = (-0.800 × 0.000 × 0.20) + (-0.200 × 0.000 × 0.80)

Weighted error sum = 0.000000

Sigmoid'(x) = 0.0000 × 1.0000 = 0.0000

Gradient = 0.000000 × 0.0000 = 0.000000

Hidden Layer 2 Gradients

H2-1:

Propagated from H3 layer (simplified)

Weighted error sum = 0.000000

Sigmoid'(x) = 0.0000

Gradient = 0.000000

H2-2:

Propagated from H3 layer (simplified)

Weighted error sum = 0.000000

Sigmoid'(x) = 0.0000

Gradient = 0.000000

H2-3:

Propagated from H3 layer (simplified)

Weighted error sum = 0.000000

Sigmoid'(x) = 0.0000

Gradient = 0.000000

Hidden Layer 1 Gradients

H1-1:

Propagated from H2 layer (simplified)

Sigmoid'(x) = 0.0000

Simplified Gradient = 0.000000

H1-2:

Propagated from H2 layer (simplified)

Sigmoid'(x) = 0.0000

Simplified Gradient = 0.000000

H1-3:

Propagated from H2 layer (simplified)

Sigmoid'(x) = 0.0000

Simplified Gradient = 0.000000

H1-4:

Propagated from H2 layer (simplified)

Sigmoid'(x) = 0.0000

Simplified Gradient = 0.000000

Weight Update Examples

Hidden Layer 3 → Output Weight Update:

For weight H3-1 → Output-1:

Current weight: 0.6000

Update = learning_rate × output_gradient × h3_activation

Update = 0.10 × 0.000000 × 0.0000

Update = 0.000000

New weight = 0.6000 - 0.000000

Backpropagation Formula Summary

Chain Rule Application:

∂Loss/∂weight = ∂Loss/∂output × ∂output/∂sum × ∂sum/∂weight

Where:

• ∂Loss/∂output = output - target (for MSE)

• ∂output/∂sum = σ'(sum) = output × (1 - output)

• ∂sum/∂weight = previous_layer_activation

Weight Update: w_new = w_old - α × gradient

Where α = learning rate

Understanding the Network

How this deep neural network works

Network Architecture:

  • Input Layer: Three nodes that receive input values (blue circles)
  • Hidden Layer 1: Four nodes that process inputs from the input layer (purple circles)
  • Hidden Layer 2: Three nodes that further process information (purple circles)
  • Hidden Layer 3: Two nodes that create final feature representations (purple circles)
  • Output Layer: Two nodes that produce the final predictions (green circles)
  • Node Values: Each circle shows the current activation value after applying the sigmoid function
  • Bias Terms: Each layer (except input) has bias values that are added to the weighted sum before activation

Mathematical Formula with Bias:

Complete Neuron Calculation:

output = σ(Σ(input_i × weight_i) + bias)

Where:

  • σ = sigmoid activation function
  • input_i = value from previous layer node i
  • weight_i = connection weight from node i
  • bias = learnable bias parameter for this neuron

How Information Flows:

  • Forward Pass: Information flows from input → hidden layer 1 → hidden layer 2 → hidden layer 3 → output
  • Activation Function: Each layer uses the sigmoid function to introduce non-linearity
  • Feature Learning: Each hidden layer learns increasingly complex and abstract features
  • Deep Architecture: Five layers allow for very sophisticated hierarchical representations

Training Process:

  • The network performs a forward pass through all 5 layers to generate predictions
  • Error is calculated by comparing final outputs with target values
  • Backpropagation adjusts weights layer by layer, working backwards from output to input
  • Each layer learns different levels of abstraction during this deep learning process

Deep Learning Architecture: This 5-layer network demonstrates true deep learning! Notice how information is progressively transformed through multiple hidden layers (4→3→2 nodes), allowing the network to learn complex, hierarchical representations of the input data.

Experiment: Watch how values change across all layers as you adjust inputs. During training, observe how the deeper architecture allows for more sophisticated learning patterns compared to shallow networks.