chapter 3 psych
Chapter 4.3:
The Russian room argument
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Lines of attack
- Frame problem
Presents difficulties for the idea of representing knowledge symbolically
- Russian room argument
Challenges the syntactic assumption at the heart of the PSSH
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Frame problem
Original version (McCarthy and Hayes 1969)
How can a formal system represent the changes brought about by an action without explicitly representing all the things that the action does not bring about?
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Broader versions of frame problem
- Some theorists have argued that the frame problem poses an in principle objection to the PSSH
• (Alleged) impossibility of formalizing commonsense reasoning
• Often accompanied by emphasis on “situatedness” and “embodiment” of real cognitive agents
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Assessment?
- It is hard to know how to assess these arguments without explicit impossibility proofs
- The real test comes with the alternative models proposed
Connectionist models of knowledge representation
Embodied/situated AI
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Syntax
- Physical symbol structures are purely syntactic
The symbols do not have any intrinsic meaning
Nor do the expressions built up out of them
The operations on physical symbols are sensitive only to the “shape” of those symbols
Formal rules, like the rules of a logical calculus
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
From syntax to semantics
- One can specify a complete machine table for a TM without saying anything about what it is intended to represent (its intended interpretation)
- The machine table just specifies what the appropriate transitions are for any possible combination of inputs and states
- But if we assign meanings to the symbols then we can interpret the machine as carrying out specific calculations
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
A sample program
Q1 0 R Q2
Q1 1 0 Q1
Q2 0 1 Q3
Q2 1 R Q2
• The symbol “R” has a fixed meaning, since it is the instruction to move one square to the right
• But “0” and “1” do not mean anything
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Running the program
Q1 0 1 1 0 1 1 0
Q1 0 0 1 0 1 1 0
Q2 0 0 1 0 1 1 0
Q2 0 0 1 0 1 1 0
Q3 0 0 1 1 1 1 0
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
*
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
An interpretation function
• An interpretation function gives a semantics
• assigns objects to symbols
“1” 1
“0” punctuation mark
• makes it possible to interpret the TM as computing the function of addition
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Syntax tracking semantics
Syntax:
“n” = a string of n “1”s bounded by “0”s
“m” = a string of m “1”s bounded by “0”s
“n + m” = a string of n + m “1”s bounded by “0”s
Semantics:
“n” designates n
“m” designates m
Isomorphism
Given inputs “n” and “m” the TM outputs “n +m” just when “n + m” designates the sum of n + m
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Background to the Russian room
- The Russian room argument exploits an intuitive contrast between
The way that the outputs of a computer result from operations on strings of symbols (“1”s and “0”s)
The way that human behavior results from rational thought involving propositional attitudes
- Searle uses the CRA to argue that this contrast is fatal to the project of strong AI (idea that appropriately programmed computers might be minds)
PSSH is committed to strong AI
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
The Russian room
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
The main claims
- The Russian Room is input-output identical to a real Russian speaker
- The “internal processing” in the Russian room is purely syntactic (based on the shapes of the symbols)
- The person in the Russian room has no understanding of Russian
Therefore, what is going on in someone who really does understand Russian (or anything else) cannot be the sort of processing that takes place in the Russian room
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
What is genuine understanding?
- Clearly cannot be understood in purely behavioral terms
i.e. producing the appropriate outputs for given inputs
The CR passes the Turing Test
- Searle: “Understanding a language, or indeed having mental states at all, involves more than just having a bunch of formal symbols. It involves having an interpretation or a meaning attached to those symbols” (In Chalmers, p. 671)
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Possible responses
Reject the intuition that the CR does not understand Russian
Concede that the CR does not genuinely understand Russian, but find an alternative explanation of the lack of understanding that does not rule out strong AI
Concede that the Russian room does not genuinely understand Russian, but show how we might build up from the CR to a system that does understand Russian
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Strategy 2: system reply
- The thought experiment is set up so that the question of whether the CR understands Russian is equivalent to the question of whether the person in the CR understands Russian
- But even if we agree that the the person in the room only has a “phrase book” understanding of Russian, this is perfectly compatible with the system as a whole having genuine linguistic understanding
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020
Strategy 3: robot reply
- The input-output test is not a good criterion for genuine understanding
It is purely verbal
- A much better test of linguistic understanding is whether the CR can interact with the world appropriately
• obey instructions and commands
• name and describe objects correctly
• initiate conversations in a relevant manner
Cognitive Science José Luis Bermúdez / Cambridge University Press 2020