Backpropagation tutorial

Overview

What is Backpropagation?

Backpropagation is a neural network training technique. Computes how much every network weight contributed to an error, then updates it. Its central idea is summarized by ∂L/∂W⁽ˡ⁾ = δ⁽ˡ⁾(a⁽ˡ⁻¹⁾)ᵀ.

This guide connects theory to practice. You will trace the input, intermediate process, and output; decode the notation; run a dependency-light implementation; then repeat the workflow with PyTorch autograd and evaluate the result.

Preparation

Prerequisites and tools

You do not need an advanced software stack. Start with a recent Python environment and the fundamentals below, then install only the packages used by the practical example.

• Comfort with Python functions, NumPy arrays, and basic descriptive statistics.
• A clear distinction between training data, validation data, and untouched test data.
• Familiarity with features, targets, preprocessing, and task-appropriate evaluation metrics.

Process

How Backpropagation works

1. 1

   Input

   A forward-pass prediction, its target, and the network’s intermediate activations.

2. 2

   Prepare and configure

   Check shapes, value ranges, ordering assumptions, missing values, and the parameters that control the algorithm’s behavior.

3. 3

   Algorithm process

   Propagate loss gradients backward through each layer with the chain rule.

4. 4

   Output

   Updated weights that make the next prediction slightly more accurate.

5. 5

   Validate

   Track gradient norms by layer to detect vanishing, exploding, or disconnected gradients.

Core concept

Formula and intuition

The notation captures the main operation or complexity statement behind Backpropagation. Read it together with the symbol key above and the step-by-step process in this guide.

Applications

When to use Backpropagation

Training neural networks for vision, language, audio, and forecasting.

Implementation

Backpropagation from scratch in Python

This dependency-light example emphasizes the algorithm’s mechanics so each important step remains visible.

import numpy as np

rng = np.random.default_rng(0)
X = rng.normal(size=(128, 2))
y = (X[:, 0] * X[:, 1] > 0).astype(float)[:, None]
W1 = rng.normal(0, .3, (2, 4))
W2 = rng.normal(0, .3, (4, 1))
learning_rate = .25

def sigmoid(z):
    return 1 / (1 + np.exp(-z))

for _ in range(600):
    hidden = sigmoid(X @ W1)          # forward pass
    prediction = sigmoid(hidden @ W2)
    output_gradient = (prediction - y) * prediction * (1 - prediction)
    hidden_gradient = (output_gradient @ W2.T) * hidden * (1 - hidden)
    W2 -= learning_rate * hidden.T @ output_gradient / len(X)
    W1 -= learning_rate * X.T @ hidden_gradient / len(X)

Practical tutorial

Build Backpropagation with PyTorch autograd

PyTorch autograd records tensor operations and applies the chain rule to populate every parameter gradient.

1. Prepare the data.A forward-pass prediction, its target, and the network’s intermediate activations. Validate its shape, type, range, and ordering before training or execution.
2. Configure the algorithm.Begin with explicit, conservative parameters and a fixed random seed whenever the library supports one.
3. Fit or execute.Propagate loss gradients backward through each layer with the chain rule.
4. Inspect the result.Updated weights that make the next prediction slightly more accurate. Then apply the evaluation checks in the next section.

import torch
from torch import nn

torch.manual_seed(7)
X = torch.randn(64, 4)
y = torch.randn(64, 1)
model = nn.Sequential(nn.Linear(4, 8), nn.Tanh(), nn.Linear(8, 1))
optimizer = torch.optim.SGD(model.parameters(), lr=0.05)

optimizer.zero_grad()
prediction = model(X)
loss = nn.functional.mse_loss(prediction, y)
loss.backward()
for name, parameter in model.named_parameters():
    print(name, "gradient norm:", parameter.grad.norm().item())
optimizer.step()
print("loss before update:", loss.item())

API details and version-specific options: official PyTorch autograd reference →

Evaluation

How to evaluate the result

A successful run is not enough. Evaluate the output against the intended use, compare it with a simple baseline, and preserve a genuinely unseen test case whenever the task involves learned parameters.

• Track gradient norms by layer to detect vanishing, exploding, or disconnected gradients.
• Use a numerical gradient check for custom operations on small test inputs.

Common mistakes

Pitfalls and how to avoid them

These failure modes are common in tutorials and production systems. Treat them as review questions, not just after-the-fact debugging advice.

• Forgetting optimizer.zero_grad accumulates gradients across steps.
• In-place operations can invalidate values needed for backward computation.
• Calling detach or converting tensors to NumPy breaks the computation graph.

Project checklist

Before using Backpropagation in a project

• ✓Define the prediction or discovery objective before selecting the algorithm.
• ✓Split data before fitting preprocessing and tune only inside cross-validation.
• ✓Record data versions, random seeds, features, hyperparameters, and evaluation metrics.
• ✓Compare against a simple baseline and inspect errors by meaningful subgroups.
• ✓Monitor input drift and real-world performance after deployment.

Further reading

Official documentation and next steps

This guide is an educational introduction, not a substitute for domain validation. For consequential applications, review the source documentation, test against representative data, and involve a subject-matter expert.

Continue learning

Related machine learning guides

• Linear RegressionFits a straight line that minimizes squared prediction errors.
• Logistic RegressionMaps a weighted feature score to a class probability.
• Decision TreeSplits the feature space into rule-based regions.