Lecture 11 - Symbol Grounding and Categories

Question 1

What is cognition?

Answer 1

Thinking
Doing-Capacity (How? Why?)
Cognitive vs. “Vegetative” function
Explicit vs. Implicit cognition
-Explicit cognition kind of like conscious processes
-Implicit cognition kind of like unconscious processes
Easy vs. Hard problem (doing vs. feeling)
(Other-Minds problem): How can you tell what someone is thinking… how can you tell they can feel at all?

Question 2

What is cognitive science?

Answer 2

Cognitive science is not brain science, but rather trying to figure out what organisms and devices that can do the kinds of things that normally humans can do: how do they do it

Reverse-engineering: Not trying to build a cognitive entity, but in order to give an explanation of cognition, kinda like reverse-engineering. The thing is built, and mission to figure out how it works

Question 3

What is a machine?

Answer 3

Any cause-effect system

Toaster, rat, clock, computer, tree, atom, planet…

We are machines! But what kinds of machine are we?

Turing machine = machine that can execute an algorithm
Universal turing machine = can execute ANY algorithm, any set of rules

Not every machine is a turing machine

Question 4

What is computation?

Answer 4

Computation = Symbol manipulation
Symbol=any object

1. Rule-based = algorithms
2. Shape-based = symbols can be any shape, arbitrary
3. Implementation-independent = in order to execute computation, countless hardwares can do it, so don’t study the hardware! It’s in the software, the algorithm. Also called the software-hardware distinction.
4. Semantically interpretable = not interested in just manipulating arbitrarily shaped symbols just for the fun of it, but doing things with them until finally you come up with a result, and the result is useful to some people

Mathematics = all syntax! Symbol manipulation (manipulating arbitrary shapes on the basis of rules)

Can’t talk about symbol grounding without language
Syntax/Semantics
* The cat is on the mat
* 貓在墊子上。

Question 5

Church/Turing Thesis
Weak and Strong

Answer 5

Alan Turing … helped win the second world war and co-invented the computer
-trying to formulize what computation was

Weak thesis: The turing machine, or other equivalent, are what it is. Just a thesis, no counter-examples, nothing that a turing machine can’t execute
Strong thesis: ? you can compute and model just about anything as close as you like ex: everything you can do with virtual reality

Mathematicians compute
Church/Turing/Goedel/Kleene/Post…
Computer simulation
Virtual objects
Virtual Reality
Squiggles and Squoggles
“Almost Everything”

Question 6

Computationalism (“Strong AI”) vs. Searle

Answer 6

• “Cognition is computation”
• “Computation is implementation-independent”
• “Turing Test is Decisive”

Searle’s Periscope
-The implementation-independence of computation
“Cognition is not computation”
Refute computationalism
Chinese Room Argument:
Argument: If I memorize that program and execute that program, I still wouldn’t be understanding Chinese
Like chatgpt… it may learn, but does it understand?

Question 7

Turing Test Hierarchy

Answer 7

• (t1) (toy) = not a turing program… like a chess playing program, scene describing program… (1) not part of the hierarchy
• T2 (verbal) : nothing can pass T2 unless it also passes T3. Ex: computer
• T3 (verbal + robotic): Robot in order to interact the referent of a word, have to move
  sensory-motor capacities of a robot to ground symbols. Ex: a robot that can talk and move (sensorimotor grounding)
• T4 (verbal + roboti + neuro). Ex: A human? Nervous sytem? Doesn’t really exist yet (idk)?

Question 8

Chinese dictionary
(Symbol grounding problem)

Answer 8

If have to learn mandarin with a mandarin only dictionary. A symbol will tell you “it means this (other symbol)”. So you search this other symbol and it leads you to more definitions like that. And you still don’t understand and it keeps going until you’ve gone through the whole dictionary and still don’t understand

All the words in the dictionary are defined by other words in the dictionary

English dictionary
Minimal grounding… ex: Around 1000 words to define all the other words…. But then it’s not the smallest number…. idk
Kernel words used not to define other words (10%)
(see image)

But how do the words in the minimal
grounding set get their meaning?
This is the symbol grounding problem

H1: Direct Sensorimotor Grounding (DSG)
Indirect Verbal Grounding (IVG)
Teacher/Learner asymmetry

H3: Feature-Detection vs.
Feature Description
-Trial and error reinforcement learning… feature-detection (detect the features of edible vs poisonous mushrooms)… direct learning
-Supervised reinforcement learning… feature description (being told (described) which are edible vs poisonous)… language
–but only works if understand the language
–the person giving the language doesn’t have to themselves understand (just like a professor doesn’t have to understand what he’s teaching to teach students)

Indirect symbol grounding
Zebra = horse + stripes

Question 9

Summary

Answer 9

Computation = implementation-independent: Algorithm does not depend on the hardware

Computationalism=everything can be expressed as an algorithm. Cogsci dominant theory

● Searleʼs Periscope:
○ Person memorized Chinese Room rulebook
○ Still no “understanding” of Mandarin
○ → We need symbol grounding, i.e. some real anchors for language concepts

Turing Test Levels:

T1 – toy ex: ELIZA
-ELIZA doesn’t have understanding of English, rule-based system that looks for patterns input-outputs

T2- verbal. Full conversation with system. Indistinguishable from human (but on a purely conversational level)

T3 – verbal + robot. Also has a body. Physical action (ex: passes me the banana)

T4 – neuronal activity. Passes if do eeg and similar brain signals to humans when answering questions

Harnad : need T3 to pass T2
-to really understand language, need to interact with something physical
-ex: minimum dictionary. Grounding for these basic words that explain every other words in the dictionary