Encoder / Decoder (Deep Learning)

In the context of deep learning, an encoder just maps high-dimensional input data into a lower-dimensional, compact representation.

Encoder: Compress input into a latent representation Decoder: Expand latent representation back into meaningful output

You’ll see “Image Encoder” for example in the case of SAM.

Examples by Context

• Autoencoder:
  • Encoder → compresses image → latent vector
  • Decoder → reconstructs image from latent vector
• Sequence-to-sequence models (like translation):
  • Encoder → reads source sentence (e.g. English)
  • Decoder → generates target sentence (e.g. French)
• VAE (Variational Autoencoder):
  • Decoder takes a sampled latent code and outputs a reconstructed image or signal
• Transformers (e.g. GPT):
  • Decoder-only: generates text from a sequence of tokens
  • Encoder-decoder (e.g. T5, BART): decoder uses encoder output + its own past outputs to generate sequences