the newell test you might say this is a soonest, vacuum airport - an unwilling calculation of sorts....

42
The Newell Test You might say this is a soonest, vacuum airport - an unwilling calculation of sorts. That would be a really weird thing to say, though.

Post on 20-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

The Newell Test

You might say this is a soonest, vacuum airport -an unwilling calculation of sorts.That would be a really weird thing to say, though.

Connectionism

Grading

Earlier Version used grades I had a problem with this If an A is human, where does any model

really belong

New rankings are less

Flexible Behavior

Computationally Universal But some things are easier than others

Real Time

2 Meanings

Model makes accurate predictions of time human processes and learning take

Slow learning in connectionist networks

Model does so in a timely manner

Adaptive

Behaviors serve needs

Globally is behavior optimal

Locally, are individual process useful

Vast Knowledge

Lots of Data Not at all well understood how the brain

does this in such a massive fashion General mechanism are a bit clearer

Dynamic

Copes with changes

Knowledge Integration

Functionality of intellectual combination

I’d like to call it Representation and Combination

Newell’s Quote

“Symbols provide distal access to knowledge-bearing structures that are located physically elsewhere within the system. The requirement for distal access is a constraint on computing systems that arises from action always being physically local, coupled with only a finite amount of knowledge being encodable within a finite volume of space, coupled with the human mind’s containing vast amounts of knowledge. Hence encoded knowledge must be spread out in space, whence it must be continually transported from where it is stored to where processing requires it. Symbols are the means that accomplish the required distal access.” (Newell, 1990, p. 427)

Dictionary

symbolnoun [C] a sign, shape or object which is used to represent something else 

What is a Symbol

Hot Debate in Cognitive Psychology Most with an opinion in the class believed the

brain could not contain symbols I found this astounding until I found out

they were using Newell’s definition Does a spring symbolize the pressure put

upon it? Many think symbols could not exist in the

brain by Newell's definition However, they will all agree the brain

represents the data…

Natural Language

Can it talk Connectionism tries to build the

capability from scratch ACT-R relies on symbolic interaction to

explain linguistic results

Consciousness

Can it inspect it’s buffers? Worse/worse

I disagree and think ACT-R is better than worse

In my opinion consciousness is the parallel matching of all productions with buffer contents

Learning

Squire Episodic Semantic Skills Priming Conditioning

Learning is in some ways fundamental to cognitive psychology so has been a big focus

Development

Does it grow ACT-R is better than worse I think

Example in the paper

Bootstrapping problem Everything is hand coded just about

Evolution

Is it susceptible to selection and variation

Brain

How do the neurons do it?

Spur to Progress

Attack of the killer bees is thwarted Makes reflection on the strengths of

other work better

Interesting compromise

Theories coming closer Thad Polk’s Idea (one of Newell’s

Grad Students) Production rule compiler Transforms production rules into

networks

Joe

Overall I found the article to be fairly clear and interesting. However, I was a little fuzzy on the topic of ACT-RN. I'm not sure I understand how the chunks are actually represented, and how the headers interact with the chunks.

ACT-RN Declarative memories

1

3

24

header

chunks

slots

Figure 1: Type Memories

5Represents chunk name

Winner take all Network

retrieval

Ian

I'm going to work with this analogy: that an assembler is to a processor as activations and connection strengths are to neurons, and that c is to assembler as act-r is to activations and connection strengths, except there isn't the same layer of abstraction between c and asm.

OK

Er, it turns out I won't reference that last sentence at all. Here's the question I have now: how does act-r account for individual differences in cognition? It is nice that the averages of human data correspond to act-r data, but it seems that a fundamental layer of paramaterization must take place to account for individual differences.

Good question. ACT-R actually has a number of parameters that might account for individual differences including W, s, d, tau, and others less commonly changed such as the 50 ms cycle time…

The general nature of act-r, with its emphasis on non-parameterization, also leads me to a related question. It seems that much of the body of knowledge on neuropsychology has come from studies of partially impaired animals. Is there any way act-r can account for the particular ways in which an individual might deviate from the averaging of humans.

We haven’t started to talk about all the parameters yet. And yes, these parameters might account for neuropsychological deficits

One more, actually. Does anything like act-r exist for animals? As an assembler is to a processor, another assembler is to another processor, and it seems similar principles may apply.

Sure, but I’m not aware any ACT-R researchers or anyone has used production rule modeling for animals…

John?

Dana

The concept of setting criteria to which all theories of cognition should aspire to fulfill seems to be an advantageous goal that would provide both a way to compare extremely different theories and to assess their respective strengths and weaknesses. However several questions about the validity of the whole process must be overcome. Since the criteria originated from a symbolic view, are they biased from the beginning? Also, will it be possible to convince everyone that these criteria are the best criteria to evaluate cognitive theories? No matter how well defined the rules of the game, if no one wants to play, it doesn't matter.

Sure, they are biased, but I’d be happy to hear their critereon.

In general, the idea of a unified architecture is an idea that stems from production system modeling. Many do not necessarily agree that the idea is profitable, necessary, etc…

I get the feeling many connectionists just use connectionism as a tool and aren’t interested in this unified archtecture concept… Some however, would agree that many of these criterea are important

Jane

why does connectionism do better on natural languages than actr?

Because it doesn’t need to be engineered to pick up on statistics in the environment… It does so automatically…

Matt

"Among other domains for which ACT-R seems to be lacking adequate mechanisms are... emotion and motivation" (p. 20) What would such a model look like? Some theories of learning emphasise motivation. How would it interact with learning mechanisms in ACT-R?

Well perhaps motivations would be kinda like productions that fired when certain drive or emotional states were active

A general curiosity of mine, which led me to learn more about ACT-R: "Sometimes the suspicion is stated that ACT-R is a general computational system that can be programmed to do anything." (p. 19) The four limitations described help me to see that this isn't true. However, I'd like to know what makes ACT-R especially suited to creating accurate cognitive models.

Hmm… What exactly do you mean? It is suited for cognitive modeling because that is what it was designed to do from the very start

Inspired by the discussion on development (p. 26): How would along-term "metamodel," encompassing possibly long term memory, or a daily juggling of various priorities interact with ACT-R?

Once we cover the production competition I think this will be imaginable… I think perhaps high-level priorities might be juggled in a similar fashion according to their “emotional” utility

This sort of high level modeling would be fantastically complex

David

The question that still sticks with me is how the Newell test is intended to solve the listed problems Newell saw in the field, if that is in fact the intention. For example, none of the criteria address the stability or diversity of theories.

I think the idea is that by having a unified model stability is fostered and diversity is unified.