🧠 Neural
Network
Feedforward — ReLU — Training
⚡ Training
▶ Forward pass
↩ Backprop
↺ Reset
⏸ Pause
Architecture Presets
XOR
Classifier
Deep
Autoencoder
CNN-like
Layer Editor
Input nodes
2
Hidden layers
2
Nodes/layer
4
Output nodes
1
Activation Function
ReLU
Sigmoid
Tanh
Leaky ReLU
ELU
Softmax
Training
Learning rate
0.05
Batch size
4
Dropout
0
Speed
1×
Live Statistics
0
Epoch
1.00
Loss
0%
Accuracy
0
Parameters
Activation Function
ReLU
f(x) = max(0, x)
Used in:
Most modern networks (ResNet, VGG).
Pros:
Fast, no vanishing gradient.
Cons:
Dying ReLU problem.
Forward Pass
Signal flow
1
Input layer
Raw feature values fed in. No activation applied.
2
Hidden layers
z = Wx + b → a = activation(z). Patterns extracted.
3
Output layer
Final prediction. Loss compared to ground truth.
4
Backprop
∂Loss/∂w computed via chain rule. Weights updated: w -= lr × ∂L/∂w
Loss Curve
Weight Distribution
Concepts
Key Concepts
Weight:
Strength of connection between two neurons.
Bias:
Threshold that shifts activation.
Gradient:
Direction of steepest loss increase — we go the opposite way.
Epoch:
One full pass through the training data.
Overfitting:
Memorising training data instead of generalising.