metaencoder

Metaencoder is a term used in machine learning to describe a neural network component that encodes meta-information about data, tasks, or models into a fixed-length representation. Unlike a traditional autoencoder, which aims to reconstruct input data, a metaencoder produces a compact embedding that describes higher-level properties or descriptors of the data or task. The embedding can then condition or guide a downstream model or meta-learner.

In practice, a metaencoder may process meta-features such as dataset size, class imbalance, feature statistics, domain labels, or task descriptors in meta-learning. The output meta-embedding can be used to adapt model architecture, select hyperparameters, or inform weight generation in a hypernetwork. Typical architectures include a small feedforward network that takes a vector of meta-features and outputs the embedding, or an encoder that processes a structured representation of a task description to yield a conditioning vector. It is often trained jointly with a meta-learner or base model, with objectives that tie the embedding to performance across tasks.

Applications for metaencoders span few-shot learning, automated machine learning (AutoML), domain adaptation, and multi-task learning. The term is not standardized and may be used with varying meanings in different research contexts. Some authors refer to a similar concept as a task encoder or context encoder within meta-learning literature.

Relation to related concepts includes connections to meta-learning, hypernetworks, conditioning mechanisms such as FiLM, and broader ideas in representation learning. See also meta-learning, hypernetworks, meta-features, and context conditioning.