Transcript Learning of neural network

CHAPTER 8.
NEURAL NETWORKS: THE
NEW CONNECTIONISM
Bodrov Alexey
OUTLINE
1.
Computer Simulation and Artificial Intelligence
2.
The Computer and the Brain
3.
Symbolic and Connectionist models
4.
Neural Networks
2
1. COMPUTER SIMULATION AND
ARTIFICIAL INTELLIGENCE
1.1 BASIC CONCEPTS
 Artificial
intelligence (AI) –
field, where scientists try to device computer systems that could
accomplish the same things as humans (it is a branch of computer
science that tries to make computers smarter).
 Computer
simulation –
attempt to mimic the functions of the human (including errors and
biases).
Reasons for making computer smarter:
Computer might do marvelous things for human (for ex. assembling
machine parts)

This may clarify questions about human cognitive process and mimic
functions of a mind.

3
1.2 CAN MACHINES THINK?




Sternberg and Ben-Zeev: “Computers cannot think, although
they can sometimes be programmed to respond as if they were
thinking”.
Turing test’s scheme: If A can do x, y, and z, and B can do x, y,
and z exactly, then B must possess whatever attributes A has
that allow it to do x, y, and z.
Conclusion: Machines can think.
Searl shown: Turing’s test reduces easily to absurd.
Conclusion: There is no answer whether machine can think or
not.
Modern view: Computers don’t need to think. Their lack of
“passion” does underscore that hey are not human.
4
2. THE COMPUTER AND THE BRAIN(1)
Basis computer-as-cognitive-processor metaphor:
Function
Structure
HUMAN
S
Cognititive acts
(thinking,
problem solving,
creating, other
cognitive
processes)
Information processing
Input
Software
(programmed
operations)
R
Senses
Wetware or
nervous
system
Response
system
COMPUTER
Output
Hardware
Sensors
Printers,
(chips, relays,
(keyboard)
screens
switches)
Opponents of this metaphor claim, that viewing humans as
machines robs them of the most important aspects of humanity
(machines have no emotion and no volition).
Penner point out that metaphors are just comparisons and we need
only accept that computers and humans sufficiently similar that some
features of one can be used as a sort of pattern for other.
5
2. THE COMPUTER AND THE BRAIN(2)
Differences between brains and computers:
o Brains are very slow; computers are lightning fast
Brains are smarter and storage of information is virtually
unlimited
o
Human nervous system is incredibly more complex than even
the largest and most sophisticated of modern computers
o
The computer’s ability to retrieve flawlessly from memory and to
perform arithmetical computations rapidly and accurately far
exceeds that of humans
o
Conclusion: Brain is more like a parallel distributed
processing (PDP) computer.
6
3. SYMBOLIC AND CONNECTIONIST
MODELS
3.1 SYMBOLIC MODELS
Basic assumptions:
o
all information can be represented in symbols
o
learning is explicit
o
information processing (thinking) involves the application of
identifiable rules
Historical illustration of this model:

Newell, Shaw, and Simon Logic Theorist, which is capable of finding
proofs for theorems in symbolic logic.
 General
Problem Solver (GPS) – is designed to allow comparison between
the desired end state and the current state. The only thing it has
revealed is that far more flexible than GPS.
 SOAR
– a model of cognitive architecture. Similar to chess programs.
Modern chess programs:
Use brute force, coupled with few key strategies
“thinking procedures” with people.
Different
7
3.2 CONNECTIONISTS MODELS
Not all learning is explicit (as symbolic model claims), for
example skill in darts.
Cognition occurs in the brain not as a series of process but more
as patterns of activation (Hebb).
Assumptions of a model:

brain’s collection of neurons is like the processing units in a PDP

no central organizer or processor governs activities of neurons
Neural Network – connectionists model of “mind”.
8
4. NEURAL NETWORKS
4.1 INTRODUCTION
“A neural network is a general mathematical computing
paradigm that models the operations of biological neural
systems”
Learning of neural network:
 New connections might develop.
 Old connections might be lost.
 Probability that one unit will activate another might
change.
In cognitive research, neural networks are not physical
arrangements of actual networks of neurons !!!
9
4.2 ILLUSTRATION: NETTALK
Task: Learn to read text.
Machine learn itself by using
back-propagation rule (uses
information about the correctness
or appropriateness of its
responses to change itself so that
the response might be more
correct or more appropriate).
Result: NETtalk could read not
only studied words, but texts it
had never seen. It learned rules,
exceptions, had learned to
generalize and made some sorts of
errors that children always
10
make.
4.3 ADVANTAGES OF CONNECTIONISTS
MODELS





Neural network can make inferences without being given
specific rules for so doing.
This models allow for a fuzzier kind of logic (more peculiar
to human).
More accurately reflect the actual physiological structure
of human nervous system.
Present a functional analogy for the notion that experience
alters the brain’s wiring.
The applications of neural networks stretch well beyond
the cognitive sciences and psychology.
11
4.4 CRITICISM OF CONNECTIONISTS
MODELS




Computers don’t simulate human emotions at all.
Computer simulations don’t reveal the insight of
which human solvers are capable.
Tell very little about how the human nervous system
works (the successful functioning of connectionist
models depends on certain properties of their units
that are not properties of the human nervous system).
Too big interference (second test of paired-associate
learning).
12
QUESTIONS
Can machine think or not?
 What is the main difference between symbolic
and connectionist models?
 What are main advantages and disadvantages of
artificial neural networks?

13
THANK YOU
14