# Why Backprop Goes Backward

> Backpropagation uses backward passes to compute neural network gradients efficiently.
- Title: Why Backprop Goes Backward · Gregorygundersen | Shortify
- Summary: Backpropagation uses backward passes to compute neural network gradients efficiently. Backpropagation computes gradients by moving backward through a neural…
- Keywords: backprop, neural network, gradient descent, chain rule, computational graph, derivative, neural networks, mathematics, algorithm, forward pass, backward pass, computational efficiency
- Source: Gregorygundersen — https://gregorygundersen.com/blog/2018/04/15/backprop
- Author: Gregory Gundersen
- Read time: 1 min
- Topics: backprop, neural network, gradient descent, chain rule, computational graph, derivative, neural networks, mathematics, algorithm, forward pass, backward pass, computational efficiency
## Why Backprop Goes Backward
Backpropagation computes neural network gradients using backward passes.
## The question
Why does backpropagation use a backward pass instead of forward?
## The problem
Computing gradients in a forward pass requires repeated calculations.
## The chain rule
Gradients can be computed locally at each node using the chain rule.
## The issue
Downstream nodes depend on current node values.
## The solution
Backward pass passes messages to compute gradients efficiently.
## Why backward?
Forward pass would have quadratic runtime in nodes.
## The insight
Backward pass computes gradients in linear time.
## The credit assignment
Nodes tell upstream neighbors what they did wrong.
## The result
Backpropagation solves credit assignment efficiently.
## The key insight
Backward pass avoids repeated calculations.
## The conclusion
Backpropagation is optimal for neural network training.
## The algorithm
Backpropagation uses backward passes to compute gradients.
## The impact
Backpropagation enables modern neural networks.
## The future
Backpropagation remains the standard for gradient computation.
## Key takeaway

Backpropagation computes gradients by moving backward through a neural network to avoid quadratic computation time.