Transcript Cognitive Adequacy and `Brain-Like` Intelligence

Chapter 15. Cognitive Adequacy in BrainLike Intelligence
in Brain-Like Intelligence, Sendhoff et al.
Course: Robots Learning from Humans
Cinarel, Ceyda
Biointelligence Laboratory
School of Computer Science and Engineering
Seoul National University
http://bi.snu.ac.kr
Contents

Artificially Intelligent Systems

What Is Brain-Like Intelligence







Perception and Action
Learning and Memory
Focusing
Motivation
Neurobionics
Cognitive Adequacy
Discussion
2
Artificially Intelligent Systems
“Machines will be capable, within twenty years, of doing
any work a man can do”
inadequate architectures and algorithms

The von Neumann architecture


One main processor is responsible for all aspects of
computation
While in human brain specialized areas process different
aspects of information simultaneously
?
?
?
?
Source: https://www.ted.com/talks/nick_bostrom_what_happens_when_our_computers_get_smarter_than_we_are#t-425849
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
3
What Is Brain-Like?




We want to imitate information processing features
of biological brains
Psychophysics has gathered a wealth of data
We can only draw conclusions about how
knowledge is represented in the human brain
Incomplete list of desired features




Perception and Action
Learning and Memory
Focusing
Motivation
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
4
Perception and Action

Through evolution:



Random mutations occurred and those who were suited
to their environment survived
Multiple sensory input and effector systems
Can’t be separated

Embodied artificial intelligence
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
5
Learning and Memory

Ability to learn



without much effort
Intimately interwoven in the neural
representation

knowledge representation & learning

processing & storage of information
concept formation and category
learning
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
6
Focusing

Relevant subsets of the data


Consolidation of memory
happens when the
information has




automatically or voluntary
high emotional salience,
behavioral relevance,
been repeated many times
Planning

Choose one of many
alternative actions
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
7
Motivation


Goal is to avoid pain and to increase reward
Sub-goals can be adaptively changed depending
on whether a drive is reduced or not.

Shouldn’t run into dead ends by trying to stick to one
strategy
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
8
Neurobionics


Exact anatomical or physiological mechanisms are
not known
Copy the functional anatomy or physiology of the
biological model
Neurobionics
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
Neuroprosthetics
9
Cognitive Adequacy

Paradox of AI

Lokendra Shastri and Venkat Ajjanagadde
“AI systems that do not require more processing time to process
hard problems compared to easy ones”

Massive parallelism is implied by adequacy


Artificial Neural Networks
Logical theorems can be clustered into different
classes of difficulty based on human performance
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
10
Chapter 15. Cognitive Adequacy in BrainLike Intelligence
in Brain-Like Intelligence, Sendhoff et al.
Course: Robots Learning from Humans
Seunghwan Cho
Computing and Memory Architecture Laboratory
School of Computer Science and Engineering
Seoul National University
http://cmalab.snu.ac.kr
Contents

Cognitive Adequacy and ‘Brain-Like’ Intelligence

Measure of Cognitive Adequacy





Reaction Time
Error Rates
Perception Measures – visual illusion or ambiguity
Limitations and Criticisms
Discussion
12
Cognitive Adequacy and ‘Brain-Like’ Intelligence

An AI system or algorithm performs cognitively adequate,
if it reveals the same relative performance measures as a biological system
solving the same task.

adequateness implies massive parallelism

physical identity (A = A, but A = a) vs. phonetic identity (A = A and A = a)

Reaction time analysis leads to that
at first, a visual representation is built up and only subsequently a more
abstract phonetic representation is established.
Measure of Cognitive Adequacy

cognitive adequacy

behaviorally rather than physiologically
Simpler Problems
An AI solves
faster
An AI shows
less errors
More Difficult Problems
slower
more errors
And AI show similar illusory or ambiguous percepts as those in human
perception

We can use abundant psychophysical data accumulated so for.
14
Measure of Cognitive Adequacy
: Reaction Time

Cognitively adequate algorithms take more time to process such conditions for
which also humans need more processing time.

Kanizsa square experiment:





Kanizsa vs. non- Kanizsa;
3 objects vs. 4 objects
Image comparison is not done as a whole but with operating separately for the two
features
reaction times attributed to mechanisms of high-level information processing not to lowlevel stimulus processing
human information processing : top-down processes
15
Measure of Cognitive Adequacy
: Error Rates

For cognitively adequate algorithms are more likely to produce an error in
conditions in which also humans are more likely to perform erroneously.

Color recognition experiment:



red, blue, green black
red, blue, green, black, orange
red, blue, green, black, orange

Three-colored disk experiment:

even though the identical stimuli were used -> significantly different error rates between the two tasks
human color processing: the processing of similarity can lead to more errors

Because similar color is likely to produce similar response.

This feature of processing of color similarity avoids errors due to changing illumination

16
Measure of Cognitive Adequacy
: Perception Measures – visual illusion or ambiguity

For cognitively adequate algorithms show similar illusory or ambiguous
percepts as those in human perception

The ambiguous Necker cube experiment:

perception alternates every few seconds
Two alternative perceptions of it

brain needs to be viewed as a dynamical system
Dynamical system interpretation of perception. Left: Multistable perception Right: In case of perceiving a less
ambiguous visual scene the state trajectory approaches a limit cycle.
17
Limitations and Criticisms

cognitive adequacy is based on an assumption.
“If the program’s input/output and timing behaviors match
corresponding human behaviors, that is evidence that some of the
program’s mechanisms could also be operating in humans.”

Does Cognitive Adequacy Guarantee ’Brain-Like’ Intelligence?
Criticism of Behavioral Tests


Chinese room argument:
18
Discussion

How human like does AI really need to be?

What more can we add to this list of features?

Can we guarantee that adequate machine
intelligence is really brain like?

What kind of motivation can an AI have?

Where does creativity fit in this schema?
© 2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
19