Cog Sci Quiz 2

Turing Test (Imitation Game)

A behavioral test proposed by Alan Turing in 1950 for whether a machine is intelligent. A human interrogator talks to a machine and a human and tries to tell which is human

Passing the Turing Test

If a machine wins the Imitation Game at least 30% of the time then we should conclude it is intelligent

Who proposed the Turing Test and when

Alan Turing in 1950

Turing's 1950 prediction

In about fifty years computers would play the imitation game so well that an average interrogator would have no more than a 70 percent chance of making the right identification after five minutes

Jones and Bergen (2025) result

When prompted to adopt a humanlike persona GPT-4.5 was judged to be the human 73% of the time in 5 minute Turing Test conversations

LLMs tested by Jones and Bergen (2025)

ELIZA and GPT-4o and LLaMa-3.1-405B and GPT-4.5

The consciousness objection

Machines are not intelligent because they do not have subjective experiences (they cannot truly feel pleasure or grief)

Turing's response to the consciousness objection

We cannot be 100% certain other people are conscious but we infer it from their behavior. If a machine passes the test we would have the same evidence it is conscious so we should accept the test

The originality objection

True intelligence requires genuine originality but everything a computer does is wholly determined by its program (Analytical Engine has no pretensions to originate anything)

Turing's replies to the originality objection

Human novelty may also be wholly determined by genes and experience. Programmers often cannot predict what a machine will do and machines surprise us (e.g. AlphaGo's surprising Go moves)

Question for Turing about originality

Is surprisingness the same as originality?

Universal decision procedure

Following a fully deterministic algorithm to decide mathematical questions (asked about by mathematicians in the 1930s)

Turing Machine

A hypothetical idealized computing device that Turing used in 1936 to define what it means to follow an algorithm

Ingredients of a Turing Machine

An infinite tape of discrete cells each holding one symbol and a read/write head and an alphabet of symbols and a set of instructions (machine table/program)

Configuration (Turing Machine)

The current state and the symbol being scanned

Automatic machine (a-machine)

A Turing machine runs on its program without external input (automatic) and its behavior is fully determined by its configuration and program (determinate)

Automaticity

Running on the program without external input

Determinacy

Behavior is fully determined by the current state and symbol and the program

Church-Turing Thesis

Every effective computation (anything calculable by a clear step-by-step algorithm) can in principle be carried out by a Turing machine

Algorithm (Turing's definition)

A procedure that can be executed by a Turing machine

Computability

Can be computed by a Turing machine

Turing's impact

The Turing Machine formalized and made precise notions like algorithm and computability

Does a Turing Machine need to understand addition to do it

No. It just follows the machine table without understanding what addition is

Weak AI

The claim that the computer is a powerful tool used to study the mind and simulate cognitive processes

Strong AI

The claim that an appropriately programmed computer really is a mind that possesses real mental states

Searle's position on Strong AI

There is no computer program such that running it is sufficient for the computer to be a mind. What is missing is semantic understanding

Semantic meaning

The meaning of representations such as the objects and concepts they refer to

Searle's Chinese Room

A monolingual English speaker in a room follows an English rulebook to match incoming Chinese character strings with outgoing ones. He mimics a fluent Chinese speaker without understanding Chinese

Searle's Room analogy

The person in the room is like AI running a program and the rulebook is the computer program. The person lacks understanding of Chinese so AIs lack real understanding of the symbols they manipulate

Upshot of the Chinese Room

Running a computer algorithm only requires manipulating symbols according to rules and does not require understanding their semantic meaning

Why mental representations matter in the Chinese Room

Mental representations are about things but for the person in the room the Chinese characters are not about anything. They are just symbols

The System Reply

The person does not understand Chinese but the whole system (person and rulebook and slips of paper) does

Searle's response to the System Reply

Imagine the person memorizes the rulebook. They still would not understand Chinese so the system as a whole does not understand it

Stats student analogy

A student who memorizes the steps to compute standard deviation can follow the algorithm without understanding the concept. A computer is like this student

The Robot Reply

Attach a camera and limbs to the computer so it connects the word cat with real cats and can touch them. This should give it understanding of the word

Searle's response to the Robot Reply

Camera images and sensory inputs are just more symbols added to the rulebook. The robot still only manipulates symbols without knowing their meaning

Churchland and Churchland Objection (1990)

Searle's thought experiment relies on untutored intuition that the person does not understand and that intuition is unreliable because of the extremely slow processing speed in the room

Churchland analogy for unreliable intuition

Waving a magnet in a dark room intuitively seems not to produce light but according to electromagnetic theory it does. Similarly processing 1000 sentences at once could change our intuition

Searle's argument against Strong AI (premises)

P1 Intelligence requires semantic understanding. P2 No computer program running is sufficient for semantic understanding. C No computer program running is sufficient for intelligence

Other factors Searle's argument allows intelligence to require

Factors beyond computation such as consciousness and free will

Marr's three levels

Computational level (most abstract) and algorithmic level and implementational level (most concrete)

Where cognitive architecture sits in Marr's levels

Between the algorithmic and implementational levels

Cognitive architecture

The underlying infrastructure of an intelligent system (Langley et al. 2009). It identifies the basic operations available and how information is encoded and stored and retrieved (Pylyshyn 1999)

Cognitive architecture vs knowledge

Architecture is the aspects of an agent that are constant over time and across domains (how the mind is structured). Knowledge is the contents being stored and processed which can change

Building analogy for cognitive architecture

Permanent features (foundation and roof and room layout) are like architecture. Non-permanent features (furniture and decoration and appliances) are like knowledge

Examples of elements of cognitive architecture

What types of memory the agent has (short-term and long-term) and what kinds of representations can be stored in them

Example NOT in cognitive architecture

The contents of the agent's memory because they change over time

Pylyshyn's (1999) question

For any pattern of cognitive behavior does it reveal a property of the mental machinery (architecture) or does it arise because of the information the machinery is using (knowledge)?

Von Neumann architecture

The standard model of a computer's architecture

ACT-R

An architecture of the human mind that is compared with the Von Neumann architecture of a computer

Working memory

A type of memory that temporarily holds and manipulates a small amount of information for immediate use. It lasts 10-30 seconds without active rehearsal

Miller's Magical Number (1956)

7 plus or minus 2 items in working memory

Cowan's updated estimate (2010)

4 plus or minus 1 chunks

Fixed information hypothesis

There is a fixed amount of information we can store in working memory

Fixed slots hypothesis

There is a fixed number of slots in working memory but each unit can hold a varying amount of information. Miller's evidence supports this

Prediction of fixed information vs fixed slots

Fixed information predicts that 839 234 175 118 is harder to store than 7 3 4 9. Fixed slots predicts they can be equally easy

Chunking

A process through which small individual pieces of information are bound together to create a meaningful whole

Why 839 234 175 118 is harder to store than 7 3 4 9

We cannot easily chunk the three-digit numbers so they do not fit into the slots

Chunking letter example

N F L C I A F B I M T V is easier than a random 12 letters because it chunks into NFL and CIA and FBI and MTV

Why working memory limits are architecture and not knowledge

The limit holds no matter what information you store

Why cognitive architecture is between the algorithmic and implementational levels

Working memory limits restrict what algorithms the mind can run (such as limited parallel processing)

Modularity

The idea that parts of the mind are organized into distinct modules rather than operating as one undifferentiated system

Mental modules

Independent information-processing systems that specialize for certain functions

Who proposed modularity

Jerry Fodor (1983) who taught at Rutgers 1988-2016

Five features of a mental module

Fixed neural architecture and domain specificity and mandatory operation and limited central accessibility and informational encapsulation

Does a module need every feature perfectly

No. It must exhibit most of them to some interesting extent and some features carry more weight than others

Most important feature of modularity

Informational encapsulation

Why the Müller-Lyer illusion motivates modules

We cannot change what we see with what we believe or know which supports a separate visual module (Hypothesis 2) over visual processing mixed with central cognition (Hypothesis 1)

Hypothesis 1 of cognitive architecture

Sensory input goes through visual processing plus central cognition together to produce what we see

Hypothesis 2 of cognitive architecture

Sensory input goes through a separate visual module to produce what we see and central cognition is separate

Fixed neural architecture

A system implemented in neural circuitry that is (a) relatively limited in extent and (b) dedicated to that system and nothing else

Fixed neural architecture example (vision)

Visual perception uses a specific pathway that does not perform other functions like audition or olfaction or general reasoning

The where pathway

Visual pathway (green) for detecting motion and location

The what pathway

Visual pathway (purple) for object identification

Specialized areas in the what pathway

Face recognition (fusiform gyrus) and word recognition (visual word form area)

Domain specificity

A module only computes over a certain type of input and only performs a restricted type of computation

Domain specificity example (vision)

The visual system answers questions about color and shape and motion and whose face it is. It does not process sounds or smells or solve problems like 23+73

Mandatory operation

The computations in a module are automatic so you cannot will yourself into not processing the stimulus

Mandatory operation example

Beta movement. Color in neighboring locations creates the illusion of a moving object and you cannot stop seeing the motion once you look at the animation

Limited central accessibility

We have conscious access to the outputs of a module's computations but not to the intermediate representations and computations that produced them

Limited central accessibility example

We can recognize faces but we do not know what algorithm our mind uses to do it

Informational encapsulation

The system has limited access to information in higher cognition

Informational encapsulation example

Knowing the two Müller-Lyer lines are equal does not change your perception that the left one looks longer

Informational encapsulation vs limited central accessibility

Encapsulation is restricted information flow from central cognition to the module. Limited central accessibility is restricted information flow from the module to central cognition

Color constancy

The ability of the visual system to perceive the intrinsic color of an object as relatively stable even when the lighting intensity and spectral composition change

How color constancy shows informational encapsulation

Knowing two blocks are the same color does not eliminate the illusion that one looks darker

Checker-board illusion

An optical illusion published by Edward Adelson (MIT) in 1995 that shows the context-dependent nature of human color perception