On Cognition

Intelligence

Intelligence should be viewed as a process rather than a skill.

Components of this process:
- World model and a way to update it
  - Learning to represent the world in a non task-specific way (joint embedding architectures).
- Persistent memory
- Internal representations
  - Supervised and reinforcement learning require too many samples/trials (regularized methods, model-predictive control).
- Reasoning and planning
  - Beyond feed-forward and System 1 subconscious computation.
  - Making reasoning compatible with learning (energy-based models?).
Keynotes of this process is generalization:
- Fluidity (synthesize new programs on the fly)
- Domain independency
- Information efficiency (abstraction and compression)
To measure machine intelligence in terms of generalization, we need to control experience and priors.
- Is that really a good idea?
- Need for benchmarks antifragile to memorization.
Examinations are not a good proxy for measuring intelligence of current machines since these were designed for humans under the latent assumptions that he/she/they need to perform generalization to do well. Such latent assumptions do not hold for machines.
Does scaling really solve abstraction? Can pattern recognition lead to pattern extraction and can that lead to reasoning?
For cognition to evolve do we not need open-endedness, agency and semi-autonomous interaction with the real world?