The Symbol Grounding Problem Qi Huang Department of Computer - - PowerPoint PPT Presentation

The Symbol Grounding Problem

Qi Huang

Department of Computer Science February 3, 2020

1 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

2 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

3 / 31

Introduction

This Paper discusses about the definition and importance of symbol grounding problem. The author proposes a hybrid non-symbolic/symbolic system in which the elementary symbols are grounded in two kinds of non-symbolic representations: iconic representation and categorical representations, and higher order symbolic representations, grounded in these elementary symbols, consist of symbol strings describing category membership

• relations. The claim is, in such a proposed hybrid symbolic

system, the symbol meanings are accordingly not just parasitic on the meanings in the head of the interpreter, but intrinsic to the dedicated symbol system itself.

4 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

5 / 31

Symbolism and formal symbol system

◮ Symbolism has been the prevailing view in cognitive theory

for several decades

◮ Belief: the mind is a symbol system and cognition is

symbol manipulation.

◮ Argument: Many of our behavioral capacities appear to be

symbolic, – the underlying cognitive processes that generates them must also be symbolic.

Question

But what exactly is a symbol system?

6 / 31

Symbol System

The definition of a symbol system is :

◮ Arbitrary Tokens ◮ Explicit Manipulation Rules ◮ Rules represented as tokens ◮ Syntactic manipulability ◮ Combination ◮ Compositness ◮ Systematicity

7 / 31

Symbol System

◮ Arbitrary Tokens A set of arbitrary physical tokens

(scratches on paper, holes on tape, events in a digital computer, etc)

◮ Explicit Manipulation Rules: Tokens are manipulated on

the basis of explicit rules

◮ Rules represented as tokens: Rules consist of likewise

physical tokens and string of tokens

8 / 31

Symbol System

◮ Syntactic manipulability: Rules are based purely on the

shape of the symbol tokens, i.e., it’s purely syntactic

◮ Combination Rules are about rulefully combining and

recombining symbol tokens

◮ Compositness: Token can be two kinds: primitive atomic

tokens, and composite symbol-token strings

◮ Systematicity: The syntax is semantically interpret-able: it

can be systematically assigned a meaning

◮ The symbolic system is independent of their specific

physical realizations

9 / 31

Symbol System: Example

Example

A thermostat turns on the furnace if the temperature goes below 70, and turns it off when it goes above 70.

Question

Thermostat’s behavior is rule-governed and interpretable – is it symbolic?

10 / 31

Symbolic System: Example

The answer is No!

◮ The rule is not explicitly represented: we are merely

interpreting the behavior as following a rule in our mind

◮ The application and manipulation is not purely formal

(syntactic, shape independent) – switching on and off is not)

◮ The system is not semantically interpretable what if it

fluctuates around 70? keep turning on and off? Is the written form of a languae a symbol system?

11 / 31

Pitfall of Symbolic System

Question

How can the semantic interpretation of a formal symbol system be made intrinsic to the system, rather than just parasitic on the meanings in our heads?

12 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

13 / 31

The Chinese Room Problem

Version 1:

◮ Suppose you had to learn Chinese as a second language

and the only source of information is a Chinese/Chinese dictionary – is it possible?

◮ Difficult – there is no end! ◮ Solution: need to ground Chinese learning in a first

language.

14 / 31

The Chinese Room Problem

Version 2:

◮ Suppose you had to learn Chinese as a first language, and

the only source of information you had was a Chinese/Chinese dictionary – is it possible?

◮ Even more impossible! You don’t even have a set of rules!

You don’t have knowledge to transfer to Chinese. Even if you memorize all the statistical patterns in the dictionary – can you create new phrases/new words?

15 / 31

The Chinese Room Problem

◮ If a computer could pass the Turing test in Chinese, does it

learn Chinese?

◮ Mere symbol manipulations have meanings that are not

intrinsic to the symbol systems itself: it’s parasitic on the fact that the symbols have meaning on us, in exactly the same way that the meanings of the symbols in a book are not intrinsic, but derive from the meanings in our head.

◮ Cognition cannot be just symbol manipulation!

16 / 31

Solution: grounding language to real world

We need to have a hybrid non-symbolic/symbolic system, to ground the "meaning" of our symbol system in sensory data, in real word experience.

17 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

18 / 31

A proposed symbol grounding system

◮ Learn to discriminate different objects with iconic

representation

◮ Learn to identify objects with categorical representation ◮ Both of these two representations are non-symbolic,

grounded in real world sensory data

◮ Create higher order symbolic representation with

symbol manipulation rules

19 / 31

Iconic representation

Iconic representation

internal analog transforms of the projections of distal objects

• n our sensory surfaces.

To Discriminate: simply a process of superimposing icons and registering their degree of disparity.

20 / 31

Identification

Icons must be selectively reduced to those invariant features of the sensory projection, i.e., categorical representation.

Categorical representation

invariant features of the sensory projection that will reliably distinguish a member of a category from any nonmembers. Both categorical and iconic representations are non-symbolic: just like images of objects in a camera.

21 / 31

Symbolic representation

◮ There are certain sets of elementary symbols that needs to

be grounded directly. "Zebra" = "horse" & "stripes": this grounds Zebra with a symbolic representation

◮ Result: someone who had never seen a zebra (but had

seen and learned to identify horses and stripes) could identify a zebra on first acquaintance armed with this symbolic representations!

22 / 31

Elementary set of symbols

One needs to has the grounded set of elementary symbols provided by a taxonomy of names (and the iconic and categorical representations that give content to the names and allow them to pick out the objects they identify). The rest of the symbol strings can be generated by symbol composition alone.

23 / 31

Grounding symbol system with connectionism

An issue remains – how do we find invariant features to establish categorical representations? Connectionism can serve this role!

24 / 31

Connectionism

We know connectionism better in : Neural network.

Connectionism

According to connectionism, cognition is not symbol manipulation but dynamic patterns of activity in a multi-layer network of nodes or units with weighted positive and negative

• interconnections. The pattern changes according to internal

network constraints governing how the activations and connection strengths are adjusted on the basis of new inputs. It’s all about patterns: the system learns and recognize patterns.

25 / 31

Comparison between symbolic AI and connectionism

◮ The former seems better at formal and language-like tasks,

the latter at sensory motor and learning tasks.

◮ Claim: neural network cannot finish many of the cognitive

activities.

◮ Argument: neural network fails to meet the compositeness

and systematicity criteria!

26 / 31

A complementary role for connectionism

Using neural network weights to identify invariant features of the category – higher weights, more likely this is the invariant feature.

27 / 31

Table of Contents

Introduction Symbol System The Symbol Grounding Problem Hybrid symbol grounding system Conclusion & Discussion

28 / 31

Conclusion

◮ Ungrounded symbolic system only possesses parasitic

meaning, and can not be the basis of true cognition.

◮ To enable a symbol system possesses intrinsic meaning,

there are more constraints than the syntactic ones.

◮ Semantic interpretations should also be "fixed" by the

behavioral capacity of the system, i.e., the decidedly non-arbitrary "shape" of the iconic and categorical representations connected to the grounded elementary symbols

◮ Symbolic AI and connectionism can form a hybrid symbolic

system

29 / 31

Discussion

◮ How to define the set of grounded elementary symbols? ◮ Is language the only/archetypal representation of

intelligence? Can other intelligent behavior be formulated as a symbolic system?

◮ How to identify new categories? ◮ Can the proposed grounded symbol system manipulate

and describe objects?

30 / 31

31 / 31