AI Blog

AI Blog

by Michele Laurelli

Exploding Gradient

/ɪkˈsploʊdɪŋ ˈɡreɪdiənt/
Concept
Definition

A problem where gradients become extremely large during training, causing unstable updates and divergence.

Exploding gradients occur when repeated multiplication of large derivatives (> 1) makes gradients exponentially larger. This causes massive weight updates that destabilize training. Solutions: gradient clipping, proper initialization, batch normalization.

Examples

1

RNN training divergence

2

NaN values in weights

3

Oscillating loss

Michele Laurelli - AI Research & Engineering