Hidden State

Hidden State refers to the internal, non-observable variables or representations within a system (biological or artificial) that determine its future behavior and outputs, distinct from its immediate inputs or visible outputs. In neural-networks, this typically manifests as the activations of hidden layers; in Cognitive Science, it relates to unconscious processing or working memory buffers.

Key Characteristics

  • Latency: Information is processed internally before manifesting in observable behavior.
  • Compression: High-dimensional input data is mapped to lower-dimensional internal representations.
  • Context Retention: Maintains information across time steps (e.g., in Recurrent Neural Networks or transformers via attention mechanisms).

AI Interpretability & Internal Representations

Recent research suggests that large language models develop structured internal spaces that function analogously to cognitive workspaces.

References