Cognitive automation encodes and operationalizes human skills and knowledge in computer systems. It enhances computer capabilities rather than replicating human minds, especially when requirements are narrow or precise examples are abundant. section 1.10
Its central limitation is adaptability. Automation handles situations represented in training data or explicit programming, while intelligence must face unknown situations, adapt to them, and learn from them. section 1.10
Deep learning makes cognitive automation unusually practical:
- Learns task mappings from data without requiring every rule to be explicitly programmed. section 19.2.4
- Generalizes locally to new situations resembling past data, but disruptions to the learned input-output relationship can invalidate the result. section 19.2.4
- Remains distinct from broad generalization, which includes handling unknown situations within a broad domain and adapting to radically novel conditions. section 19.2.4
Automation versus broad intelligence
A self-driving car that safely handles any situation would demonstrate broad generalization. A system trained for familiar driving conditions illustrates cognitive automation: capable within learned boundaries, but not necessarily prepared for situations its creators did not anticipate. section 19.2.4
