Project index
These models capture a host of features in human (mainly children and young adults)
behavior in doing "laboratory" type tasks ... in a kind of "pin the donkey" task ...
and to a lesser extent, mathematics.
For those interested in goodness of fit results with "real" data, we refer them to
the tasks in CAPITAL LETTERS on this following list:
STRATEGY ABSTRACTOR: MECHANISM CLONING IN A MULTIPLE MEMORY SCHEME
Strategy Abstractor: Original/Native Version
SPREADING INHIBITION: BASIC MODEL --- ROUND 1
SPREADING INHIBITION: ENHANCED MODEL --- ROUND 2
FeedForward Network (Instruction Learning/Following) Models: General
FeedForward Network: Architecture
FeedForward Network: EnCoding
FeedForward Network: Training
Enhanced Hamming Network Models
Rule Based and Logic Processing Models
Collected Results --- Save for techniques (.../collected_results/0918).
* STRATEGY ABSTRACTOR .... Partial Clone Model
GROUP II: Partial Clone Model
PCM is a SA model which adapts its component activations in part from learned values
(weights) in the "components model" (see Project Index link at top of this page);
its additional values are derived from standard SA mechanisms (ref: grant application).
A point to be noticed, the origin of these weights is deemed "not critical."
Their "statistics" is important, however! And, in off-line studies we "captured"
these in a feasible exercise.
What is important is the structure AND functioning of the LAYERS in both
the SA_C and the components model is the same ... so that the same learning
model can be applied to either model --- Note: this is the ONLY point in the
model where significant learning occurs.
Further discussion on using "borrowed" weights has several dimensions:
1) originally the components model had great difficulties with accuracy and
its later successes in this regard came only after SA_C successes (which
showed a new way to analyse the data);
2) the components model does not depict intermediate placements of objects,
needed in developing realistic descriptions of the overall performance
--- ergo: model expansion is needed ... with (it might be added) some newer
thoughts on training;
3) explorations in multiple mechanisms can be justified since many features of
the studied behaviors have not been addressed, the research having
subordinated these to desires to focus on key strategy emergence and, more
recently, perhaps, accuracy --- many of the required mechanisms have been
worked out for MM models: in the SA, in FeedForward Layered Neural Nets,
Spreading Inhibition, and Hamming-like Nets.
For GOF Model-Data results .... Hit this link: SA Strategy & Accuracy Results
* STRATEGY ABSTRACTOR .... Original Models
GROUP I: Native Versions
The SA-NV model is a self-contained strategy development and accuracy results
generating model.
Hithertoo, the model has been used primarily to account for behaviors/features like:
1) carrying out instructions in the wrong sequence; (sequence/order errors);
2) ability to query a subject about the instructions within a trial, e.g., what is
the object in instruction 3 ... what target "goes with" the object, pencil, etc.
3) using a variable number of instructions within trials;
4) accounting for subjects' not consistently using the best strategy known by them;
5) integrating "central" associative schemes with afferent systems that determine
more precisely where objects are placed and which potentially allow errors to
emerge outside the locale of the central associations
6) setting the stage for the subjects to generate fuzzy rule systems interpreting
their strategy behaviors in terms of "orienting" and "non-orienting" strategies,
and others (e.g. OTR + Sequence), as well as just O, OT, OTR.
* SPREADING INHIBITION I --- Basic Model (Round 1)
The Spreading Inhibition (SIM) model's approach to defining strategies is explained
in Reilly et al.'s 1996 paper (see Project Index --- section/reference list
not yet built). We do not repeat those results here, and concentrate fully on
Accuracies derived from these (pre-defined) strategies.
Our first report is what we call the Basic Model, which passes all its Chi-Square
GOF tests. Our second improves on the first.
SIM admits a variety of other observations, which we do not report on here, e.g.,
identifying wrongly delivered objects and wrongly address targets and relations;
primacy and recency effects; information retrieval possibilities; internal as well
as external memory strategies (rehearsal schedules); and parsing mechanism based
schedules of inhibition spread.
For GOF Model-Data results Basic Model: Hit this link: SI Accuracy Results --- Round 1
* SPREADING INHIBITION II --- Model Round 2
This interpretation of SI model outputs provides potential for an improved performance
over and above the results seen in Spreading Inhibition I --- Basic Model. The model
materializes a portion of Strat 5 ... It can markedly improve scores --- we show "some"
improvement -- click on link below --- Round 2 results include a Round 1 comparison.
Theoretical justification are being developed for the assumptions here ... in the present
compare, we did it "easy" by changing only Strat 3 acto the "new plan." Our solution, were
we to alter ALL Strats probably would be easier theoretically ... We could have a Round 3 --
For GOF Model-Data results Basic Model: Hit this link: SI Accuracy Results --- Round 2
* FEEDFORWARD (Layered) MODELS
Feedforward models provide a rich substrate for considerations, most of which will
be covered in papers in preparation.
We divide the studies done so far into three main areas/issues:
1) NETWORK DIMENSIONALITY: single layer, dual layer systems and other schemes
2) ENCODINGS: to include an entire trial, elements of instructions, etc.
3) TRAINING: to include methods, number of exposures, etc.
There dimensions are inter-locked, though a series of inquiries may keep one or
two of them fixed, while exploring the third as an independent feature.
Themes have included, with one or two only beginning to be investigated:
1) learning with few trials .. while achieving a given level of accuracy
2) learning with many trials .. achieving a maximum level of accuracy
3) training schedules for case 2)
4) erroneous responses, as a consequence of the input (driving force)
5) erroneous responses, in terms of output effects, to include synonmity
6) response capability in relation to corruptions and deletions in input
7) response capability in relation to corruptions and deletions in weights
A new theme has appeared of late, in response to queries from John Drake: specifically,
a FFW model that captures input-output correspondences in the fashion an extended SA
model might. The extension would be due in part to having a NN scheme which produces
intermediate movements prior to final deliveries, but the key matter is to model
appropriately chosen inputs and outputs so as to capture observed/observable behavior
of real subjects.
NETWORK DIMENSIONALITY: single layer and dual layer systems
Both single and two layer models have been discussed, with a cameo role played by
multiple layer systems. Still other experiments have involved development of
"off-line" NNs which capture some portion the aural and vision data available to
the subject.
ENCODINGS: Minimum Coding, Min. Category Coding, Amplified Codings (2 to date)
A potential exists for capturing some of the same fundamental effects by dealing
directly with coding schemes. Hithertoo, we have accentuated the following schemes:
1) minimal (trial-level) encoding
2) minimal category coding
3) augmented minimal category coding ("modest" augmenting)
4) augmented minimal category coding ("extensive" augmenting)
Interpretations with these coding schemes allow for assumptions such as coding
for properties such as side, color, position (in the viewing field), alternative
names, etc.
TRAINING: BackPropagation and others (for single layer cases)
Different methods have not proved to be a major concern. Accent has been on
speed and accuracy.
Training schedules can be important in achieving maximum learning. Investigable
is the potential these might have in shorter-time learning.
* ENHANCED HAMMING NETWORK MODELS
The EHN model was developed with these goals in mind:
1) capturing salient features of the "Relations" task (known in the vernacular as
the "Ghost" task)
2) to model a form of NN "central control" potentially relatable to Baddeley's notion
by this (same) name
3) to provide ability to handle variable numbers of instructions and to utilize info
on number of instructions as part of the input in helping to decided stategy
4) to provide graphical display of all action, including intermediate placements of
objects
5) more
* RULE BASED and LOGIC MODELS
Explorations of RB and LP mechanisms have been enjoined to try to capture schemes of
these types which can possible be rendered in NN form. These explorations may be
viewed as feasibility studies whereby we learn what is required to master the
tasks at hand.
Previous work with Prof. Y. Hayashi, Dr. J. Buckley, and Dr. P. V. Krishnamraju
has explored combined rule and NN systems. We may link this new work to that;
the earlier work has provided insights. Fuzzy rule bases have played a key role.
Learning, e.g., advanced arithmetics and function processing, in fuzzy NNs has
been involved.
A full model of the tasks represented by the components and strategy abstractor
models has been developed in logic programming. It includes a simple scheme to
learn the key weights of the components model, providing a mechanism whereby the
SA model can utilized its own weights, or, mutatis mutandi, borrow yet again.
In point of fact, the SA cloning may be cloning of a fuzzy logic rule base model.
Another rule-based effort seeks to develop a dialog between user and system. By input'g
of appropriate information, the system gives a "diagnostic" in the form of a "probability
of error" given the input (on strategy, etc.)