3

Humans have an amazing capacity for pattern recognition. I don't understand Searle's assertion "it seems to me quite obvious in the example that I do not understand a word of the Chinese stories." Of course, he wouldn't understand after only a few iterations. But, as he continues to transcribe wouldn't he naturally recognize more and more patterns and correlations and, given infinite time, eventually learn Chinese?
Is there something that would prevent him from learning it?
Is it possible that Searle could learn to understand Chinese, but not the computer?

Summary of Searl's Thought Experiment

Suppose a computer passes the Turing Test in Chinese. Now, suppose I am alone in a room and perform the same computational tasks as the computer, but manually. After the manual program has run, the person who was having a conversation with the "computer" (me) assumes they were corresponding with a person who speaks Chinese. However, I still don't know Chinese. Therefore the Computer also doesn't know Chinese.

Click for full of John Searle's "Minds, Brains, and Programs."

  • Related: in 2010, MIT researches managed to decipher the ancient language Ugaritic. However, this was done with a lot of assumptions, mostly about the language being morphologically and syntactically close to other known languages (other Semitic languages in this case). So, I'm not sure whether computers can already decipher a language of which nothing is known. – user2953 Jul 7 '17 at 6:52
  • Thanks for the recommendation. | Yes, but in the context of hard AI, there would be a learning or self-programming mechanism. It seems to me his thought experiment falls apart because the computer could add lexemes, semantics, etc as to a database (equivalent, but superior to his memory) and additionally build up a pattern database comparable to modern image matching databases (equivalent, but inferior to his pattern recognition capability). – Rubellite Fae Jul 7 '17 at 7:05
4

The argument is against the validity of the Turing test as a sufficient sign of intelligence. In a comment, you say:

Perhaps I should reword the question to Given enough time, data, & quality programming what makes us think the computer proper can't learn via pattern recognition & analysis as a human could?

Searle's argument doesn't dispute this, it provides a (hypothetical) system which arguably passes the Turing test without being intelligent. In the case you describe, the system already responds as if it knows Chinese before it learns it, so the 'knowing Chinese' later version of the system is indistinguishable from the 'not-knowing Chinese', un-understanding earlier version of the system.

Since it's possible to construct a system which does pass the Turing test but is not intelligent shows that a system that passes the Turing test needn't be intelligent. The Chinese room argument doesn't aim to prove that it's impossible to build an intelligent system, or that a system that passes the Turing test is necessarily unintelligent.

5

Searle isn't suggesting that he can't learn to understand Chinese. He's asserting that, just because a computer can answer questions like a human, does not mean that the computer understands language.

In his thought experiment, he's just like a computer when it comes to Chinese because he doesn't understand it. You shouldn't get hung up on it being Chinese or English or Voynich script or whatever. The key piece to his thought experiment is that he doesn't understand the symbols and is just applying an algorithm.

Now, Searle is using a bit of smoke and mirrors in his thought experiment. By having him in both the understanding and non-understanding process, he's making you compare the intelligent him from the dumb him. His sleight of hand is in distracting you from realising just how astonishing his paper algorithm would need to be to appear intelligent.

For example, say the questions started:

  • What is your mother's name?
  • How many children has she?
  • How old are they?

Firstly, note that the algorithm can't just answer questions, it needs context. It can't just look up How old are they? and produce an answer. It can't even just go back to the previous question as it won't know who she is. So it will need to have maintained contextual information through the whole conversation, in general, based on questions, answers and in-built knowledge. And not just trivial context, it would have needed to equate mother to she and work out that how old refers to the children. So, in order to appear intelligent for a simple exchange like this, the algorithm would need to be highly sophisticated; dealing with complex language issues, complex data interactions and state management.

So, it may be fair to argue that the 'hardware' the algorithm runs on i.e. Searle doesn't have understanding. But I'd argue that the combination of the algorithm, the execution of the algorithm and the data could.

  • Sure, I understand the argument. Perhaps I should reword the question to Given enough time, data, & quality programming what makes us think the computer proper can't learn via pattern recognition & analysis as a human could? – Rubellite Fae Jul 7 '17 at 15:47
  • @RubelliteFae That's a substantially different question from your original, not a rewording. – Chris Sunami Jul 7 '17 at 16:55
  • As he continues to transcribe wouldn't he naturally recognize more and more patterns and correlations and, given infinite time, eventually learn Chinese? The only difference is "he" and "computer." But I was taking these as equivalent. – Rubellite Fae Jul 7 '17 at 16:57
1

To answer the question as actually phrased: NO. Consider what the person in the room actually sees: Fragments of Chinese (in and out) plus (presumably) various internal indexes and directions for updating the internal system state. At what point is any of this actually connected to the outside world? Never, unless the system contains extra notes that the person can read - but I think we have to exclude that, or at least I am sure Searle would exclude that [analogy: would a 'grandmother neuron' be explicitly labelled with the word 'grandmother'??].

People learn language by connecting words with the world - we learn what a ball is when someone says 'ball' while showing us a ball, we learn what 'large' is when someone says 'large' while (e.g) waving their arms expansively, and so on. In theory, the person in the room MIGHT learn how to issue Chinese symbol sequences in response to other sequences, in a sort of statistical pattern-matching sort of way - but there would be no meaning for the person, because none of it would be grounded.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.