[ACT-R-users] Spreading Activation in ACT-R

[Kelley, Troy (Civ,ARL/HRED)]

More generally, whenever a composite chunk is competing with
noncomposite chunks (e.g. words vs. letters, phrases vs. words, objects
vs. features, collections vs. individuals), without carryover of source
activation, the elements of the composite chunk will need to be retained
in the focus of attention to activate the composite chunk given the
current implementation of spreading activation (and assuming retrievals
are not specific to the composite type). So long as composite chunks are
limited to a few slots and values, retaining the noncomposite chunks in
the focus of attention should work to activate the composite chunk.
However, carryover source activation offers an alternative to this
approach.

 

I think a composite chunk would be the best way to handle the problem in
the current implementation.  And, the problem of a composite chunk
competing with a noncomposite chunk would seem to go away if the
activation levels were higher for composite chunks.  For example, after
seeing the group of words, "kicked the bucket" repeatedly, act-r could
create a new chunk for the entire phrase out of separate chunks for each
word, based on activation levels.  This could be done at the production
system level, but it would also be interesting to have this as an
architectural feature as well.  Anyway, at that point, you could assume
a higher level of activation for the composite chunk, "kicked the
bucket", as compared to a lower less occurring word like,
"gubernatorial".  Furthermore, spreading activation can be thought of as
a computational approach to context.  So, reading a paragraph about
someone's impending demise would spread activation to phrases like
"kicked the bucket" rather than "kicked the can".

 

All that being said, I like the idea of carryover of source activation.

 

In ACT-R, the activation of a DM chunk is a combination of the base
level activation and the source activation spread from buffers in the
focus of attention. The activation of all DM chunks (consistent with the
retrieval template) is computed (in parallel) at the time of a retrieval
and the most highly activated chunk is retrieved (subject to noise). It
has recently been brought to my attention that there is no carryover of
source activation from DM retrieval to DM retrieval. The source
activation is recomputed on each retrieval given the contents of the
buffers at the time of the retrieval. On the assumption that base level
activations reach asymptote for frequently used DM chunks (following the
power law of learning), it is unclear how base level activation alone
can account for recency effects of DM chunks which have reached
asymptote, since changes in base level will be "in the noise". Something
like carryover source activation would appear to be needed.

 

The asymptotic nature of the activation equation can be a little
problematic given the task.  I have been looking at developing an
architecture that has various levels of activation or decay for
different memory chunks within the same model.  Basically, some memories
activate or decay much faster than others depending on the type of
memory they are (i.e. perceptual, low level memories decay faster than
high level conceptual memories).  This allows some more control over
when the memory reaches asymptotic levels.  I have also been looking at
using a "return-to-zero" function, so that once a memory asymptotes, it
can be "reset" to zero depending on the situation.  This would simulate
the effect of a low-level neuron which fires, goes through a refractory
period (similar to the asymptotic part of the activation function), then
eventually returns to "zero" where it can then fire again starting from
zero.  I am not saying that all memory functions this way, or that
memories are the same as neurons firing, but it is a useful function to
have when modeling memory. 

 

Troy Kelley

U.S. Army Research Laboratory

Human Research and Engineering Directorate

AMSRD-ARL-HR-SE, APG, MD

21005-5425

Voice: 410-278-5859

email: tkelley at arl.army.mil

 

  _____  

From: act-r-users-bounces at act-r.psy.cmu.edu
[mailto:act-r-users-bounces at act-r.psy.cmu.edu] On Behalf Of
Jerry.Ball at mesa.afmc.af.mil
Sent: Tuesday, October 25, 2005 2:33 PM
To: act-r-users+ at andrew.cmu.edu
Subject: [ACT-R-users] Spreading Activation in ACT-R

 

In ACT-R, the activation of a DM chunk is a combination of the base
level activation and the source activation spread from buffers in the
focus of attention. The activation of all DM chunks (consistent with the
retrieval template) is computed (in parallel) at the time of a retrieval
and the most highly activated chunk is retrieved (subject to noise). It
has recently been brought to my attention that there is no carryover of
source activation from DM retrieval to DM retrieval. The source
activation is recomputed on each retrieval given the contents of the
buffers at the time of the retrieval. On the assumption that base level
activations reach asymptote for frequently used DM chunks (following the
power law of learning), it is unclear how base level activation alone
can account for recency effects of DM chunks which have reached
asymptote, since changes in base level will be "in the noise". Something
like carryover source activation would appear to be needed.

 

As a practical example of where carryover source activation would be
useful, considering the recognition of idiomatic expressions like
"kicked the bucket". At the processing of the word "bucket", recognition
of the idiom "kicked the bucket" instead of just the word "bucket"
requires activation from both "kicked" and "bucket". With carryover
source activation, the activation of "kicked the bucket" by "kicked"
would carryover to the activation of "kicked the bucket" by "bucket".
Without carryover source activation, it will be necessary for "kicked"
to remain in the focus of attention along with "bucket" in order to
spread source activation to "kicked the bucket". 

 

More generally, whenever a composite chunk is competing with
noncomposite chunks (e.g. words vs. letters, phrases vs. words, objects
vs. features, collections vs. individuals), without carryover of source
activation, the elements of the composite chunk will need to be retained
in the focus of attention to activate the composite chunk given the
current implementation of spreading activation (and assuming retrievals
are not specific to the composite type). So long as composite chunks are
limited to a few slots and values, retaining the noncomposite chunks in
the focus of attention should work to activate the composite chunk.
However, carryover source activation offers an alternative to this
approach.

 

Speculating on other potential uses of carryover source activation,
whereas long-term learning would presumably be reflected in base level
activations, short term learning effects might involve an interaction of
base level and carryover source activation. This could provide an
explanation for spacing effects in learning and the discontinuity in the
rate of decay that is often found across short and longer time frames.

 

Even more speculatively, carryover source activation might be viewed as
residual neural activity akin to Grossberg's notion of resonance,
whereas base level activation results from long term changes in neural
potentiation.

 

Jerry

 

  

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/act-r-users/attachments/20051025/47582832/attachment.html>