Glossary · Deep learning
What is Knowledge distillation? Definition and meaning
Training a compact model to match the outputs of a larger model, transferring much of its capability.
Definition
Knowledge distillation: Training a compact model to match the outputs of a larger model, transferring much of its capability.
In short: A small student learns from a big teacher.
When it is used
Making models smaller and faster for deployment.
Where it fits
Knowledge distillation belongs to Building blocks in the Deep learning topic. These are the other terms in the same group:
- Attention mechanism: A mechanism that computes weighted combinations of inputs, with weights based on relevance between queries and keys.
- Backpropagation: An algorithm that uses the chain rule to compute gradients of the loss with respect to every weight.
- Normalization layers: Layers that rescale activations to a standard range, stabilising and accelerating training.
- Dropout: A regularisation technique that zeroes random activations during training to prevent co-adaptation.
- Optimizers: Algorithms that update model parameters from gradients, such as SGD, momentum, Adam and AdamW.
- Word embeddings: Learned vector representations of words where similar words have similar vectors, as in Word2Vec and GloVe.
- U-Net: A convolutional encoder-decoder with skip connections that preserves fine detail for pixel-level predictions.
- Pruning: Removing weights, channels or layers that contribute little, to shrink and speed up a network.
Learn more
The full guide to Knowledge distillation explains how it works, when to use it and how it compares with the alternatives, with examples. The glossary lists every term on the site from A to Z.