Skip to content
Tutorials Center
My learning

Topics

Lesson 34 of 43 ยท Artificial Intelligence and Cognition

Artificial Intelligence and Cognition: Chapter Review

Review chapter 7: the imitation game, the Chinese room, search versus learning, CNNs and the ventral stream, dopamine as prediction error, and language models, with a quiz.

Chapter 7 examined what building machines has taught cognitive science. This review collects the main arguments and results, and gives you a quiz to check your understanding. The chapter's moral is double: machines are powerful tools for testing ideas about minds, and similar behavior never settles whether the mechanism is similar.

Three cards summarize history, brains and networks, and language models.
Chapter 7 in one picture.

What you will learn

  • How to place each milestone of AI in a line of argument.
  • Which findings link learning machines to brain data.
  • What the open questions about language models are.

Recap table

Idea What it says Key names Limit or caution
Imitation game Judge intelligence by conversational behavior Turing 1950 Can be passed by tricks; Eliza effect
ELIZA Pattern matching can seem to understand Weizenbaum 1966 No understanding at all
Chinese room Symbol manipulation is not enough for understanding Searle 1980 Systems, robot and brain-simulator replies
Deep Blue Brute-force search beats the champion 1997 Not human-like play
AlphaGo Deep learning plus search beats a Go champion 2016 Trained with huge compute and data
Moravec's paradox Perception and movement are harder than logic Moravec; Dreyfus Informal generalization
CNNs and cortex Layers resemble stages of the ventral stream Hubel and Wiesel; Yamins and DiCarlo Adversarial fragility, data hunger
TD learning Dopamine signals reward prediction error Sutton; Schultz, Dayan, Montague Dopamine does more than this
Language models Next-token predictors with surprising abilities Vaswani et al. 2017; Bender et al. 2021 Understanding is contested

Questions to carry forward

  1. When is it justified to say a machine "understands"?
  2. What would count as evidence that a mechanism, and not just behavior, matches the brain?
  3. Which parts of human learning are missing in current machines: efficiency, grounding, embodiment, motivation?

Chapter 8 turns to the hardest version of these questions: consciousness.

Check your understanding

Quick check ยท 10 questions

of 10 answered You scored / 10 ยท perfect!

  1. 1. Turing's 1950 proposal replaced 'Can machines think?' with:

    The imitation game is a text conversation test.
  2. 2. ELIZA is important because it showed:

    People read understanding into ELIZA's scripted replies, the Eliza effect.
  3. 3. Which reply to the Chinese room says the whole system understands?

    The systems reply locates understanding in the entire system, not the person alone.
  4. 4. Deep Blue and AlphaGo differ mainly in that:

    Chess fell to brute-force search, while Go required learned evaluation and move selection.
  5. 5. Moravec's paradox says that, for computers:

    Skills refined by long evolution are hard to specify, while formal reasoning is comparatively easy.
  6. 6. CNNs optimized for object recognition predicted what, in the work of Yamins and DiCarlo?

    Their internal representations matched neural responses at the top of the ventral stream.
  7. 7. In temporal-difference learning, the error is:

    The TD error compares successive predictions with the reward received.
  8. 8. Schultz's monkey dopamine neurons, after learning, fired to:

    The prediction error shifted to the earliest reliable predictor.
  9. 9. An adversarial example shows that networks can be:

    Such fragility is one difference between deep networks and human vision.
  10. 10. Which is a fair statement about language models?

    The evidence is mixed and the field moves quickly.

Summary

  • AI has served as a test bed for theories of mind, with failures teaching as much as successes.
  • Search and learning solve different problems, and neither alone is how humans play.
  • Deep networks and RL link to brain data, but with clear differences.
  • Language models are a new, unsettled puzzle about understanding.

Next: The Science of Consciousness: Chapter Overview