As is, I cannot comment in any way aside from wild speculation so I will have to wait on posting the data. 

Will Learning Models ever get it right?  My response was that they already have. A learning model is a mechanistic, often mathematical, description of the inter-relation of theoretical variables designed to describe how learning occurs. That is, a Learning Model is a precise specification of the syntax of a theory. Such a model is, thus, appropriately grounded by rules that define how those theoretical variables, or constructs, relate to empirical, measurable, events.

These models, at least those coming from Associative Learning theory, already have “it” right because Associative Learning has “it” right.

Every phenomenon that is explained by a principle, exemplified by a model, is a snapshot in time of the operation of many variables. Within that snapshot, for which the model was developed, the model is correct. As more variables are added, changed, or otherwise considered, the explanation may no longer hold. That is simply the way things work. No field of science enjoys the ability to develop a model or theory that will continue to accurately account for the state of things when new variables are added to the snapshot. 

Thus, failings of models based on theories found in Associative Learning are not to be taken as failings of the general Associative Learning approach to understanding human psychology. I truly believe that any other perspective on that issue represents an immature or otherwise underdeveloped understanding of the nature of science, explanation, and the operation of the natural world in general.

The “association” and the accompanying representations and responses that are associated, are the proper level of analysis for human and animal behavior. The association has sufficient face validity to be generally understood by the lay public, and it is a sufficient construct to make connections across the fields of psychology. There is no area which does not contain some version of the idea of an “association.” The construct can be further reduced to be described and explained mechanistically, mathematically, or neurobiologically. The “Association” is an appropriate level of analysis in and of itself that can be expanded or reduced depending upon the level of explanation desired.

Associative Learning is a process. It is leverage. A screw jack will lift a car, and a floor jack in a garage will lift a car. Mechanistically they are very different, and that is necessary because they operate in different environments (an analogy to evolution operating on living beings). Nevertheless, the principle by which they operate is the same. It is “leverage.” Though too often unrecognized, Associative Learning is the “leverage” of psychology.

The “spectacular data” are those presented in the second and third experiments in this post, though those in the first experiment are interesting as well.   I’m presenting these data in the general form as I presented them in a poster, in the hallway, on the last day, on the last session, of the 17th meeting of the European Society for Cognitive Psychology.

The method is the Learning Game, described elsewhere in this blog.  In the last two experiments we used an SMIVision RED system to monitor participant’s gaze.  The Cliff’s notes are that a keypress response is attached to a spaceship.  Later, that spaceship is paired with a sensor CS.  Learning is evidenced by the emergence of the keypress response to the CS.

The motivation that I had for combining and presenting these data as I did came from some reviewer comments I received on a recent paper that I had submitted.  The question in that paper is unimportant, but what aroused my attention was that the reviewer suggested that perhaps participants were “uncertain” and then used that vague notion to explain the findings.   I found myself at a loss for a response.  In short, I suppose in my own smug way I couldn’t quite believe I needed to respond.

I began to wonder how the use of such a term became acceptable.  I then recalled a quote from Skinner:

“I accuse cognitive scientists of reviving a theory in which feelings and state of mind observed through introspection are taken as the causes of behavior rather than collateral effects of the causes.”

Skinner, 1987, Upon Further Reflection, Englewood Cliffs: NJ: Prentice Hall.

I took that statement as mostly a criticism of the methods of cognitive science, and a lack of appreciation of the structure and function of properly anchored theory development where the constructs, at least within an experimental setting, can and should have precise definitions.  I then began to look at work in my field that attempts to pursue understanding of associative learning using introspective methods I view those methods as any task that relies on self-report of an internal state that would supposedly result in behavior, as opposed to the behavior itself, such as causal/predictive learning tasks.  I became frustrated at the degree to which such theory anchoring is overlooked in my own precious field.

It is unclear to me what it is that has led us to “believe” that we must ask people questions about their beliefs/mental states and how they supposedly change as a function of our manipulations when rich inferences remain possible through the study of behavior.

Thus, in addition to presenting the data and the interesting phenomena to which they apply, I had the more pithy and high-minded goal of demonstrating just what I allude to in the previous paragraph: Complex processes are accessible without introspection.  Just exactly who I think I am is unclear””, but I felt like preaching a bit with the last sentence of the abstract.

The abstract is below.  In short, I present three experiments which demonstrate that participants learn “what, when, and where” events can occur, and that such learning obeys principles derived from associative learning. 

I’m cutting and pasting directly from the presentation, hence the gold-on-black formatting.

Nothing terribly “spectacular” about the data above, but they do demonstrate a solid “blocking” effect and show that response timing emerges with the conditioned response.  Given that the blocking group is showing a weaker response than the controls, and that the response is still well-timed, one could argue that timing even emerges before the response.  These should be interesting data for anyone who has followed Peter Balsam’s work on such issues (e.g., Balsam, P. D., Drew, M. R., & Yang, C. (2002). Timing at the start of associative learning. Learning and Motivation, 33, 141–155.).

Now to the Spectacular Data…  The next experiment was undertaken to determine if human subjects will goal-track.  That is, do they come to direct their gaze to the areas of the screen where the US will arrive.

We also monitored the participant’s pupil dilation, with the hope that pupils would come to dilate to the CS as they began to expect the US. 

Pupil dilations are not very interpretable with the present method because participants are responding during the CS, and that responding alone can cause the pupils to dilate.  Work from the lab of Harald Lachnit has shown that pupils will dilate in anticipation of responding.  With this method, all participants respond to the US, but some never learn to respond to the CS.  When I remove these non-responders and look at their pupil data, they also show the dilation response.  That is, their pupils dilate during the CS even though they are not responding at that moment.  Thus, the variable responsible (responding on the key or anticipation of more responding on the key) for the pupil dilation among the responders is ambiguous.

The method is the video game task described elsewhere where responses are first attached to spaceship USs, and then those spaceships are signaled by sensor CSs.  As participants associate these events, responding emerges to the CS before the US appears.

The design was very simple. 

•Group Latent Inhibition  6 R- |10 R+ | 10 R-

•Group Control  6 — |10 R+ | 10 R-

Group Latent inhibition received 6 presentations of a red sensor CS, while Group Control simply sat and observed the game with no events occurring.  Then, both groups received 10 trials where the red CS was paired with a spaceship that emerged from within the top left quadrant of the screen.  Following conditioning, both groups received 10 extinction trials.

Pre-Exposure

Group Latent Inhibition: First pre-exposure trial

The screen above is a heat map of the first pre-exposure trial.  What it shows is that the LI group looked at the sensor (circular area) more than the rest of the screen. Compare the pattern to that of the Controls below.

Group Control: First pre-exposure trial (No stimulus presented)

Here, the group distributes its gaze more uniformly across the screen, with some focus on the station (not shown in the background used for the heat map). 

By the last pre-exposure trial, the pattern had changed.

Group Latent Inhibition: Last pre-exposure trial

Although the sensor was looked at repeatedly more often than the rest of the screen, other hotspots emerged, thus, looking at the sensor had declined.   The control group, on the other hand (see below), was showing essentially the same pattern as on the first trial.

Group Control: Last pre-exposure trial

There is an interesting pattern to point out here.  Look closely at the last two maps above and you might notice that there appear to be more points on the left of the screen than on the right.  This trend was reliable. Participants, for whatever reason, preferred to look at the left of the screen over the right side.   The attending researcher was seated to the participant’s right, but whether that was a contributing factor is unknown.

The next task was to quantify the data presented above, the results of which are presented below.

The figure shows the seconds spent looking at the CS for each group, for each second of each trial.  Notice that the control group (solid circles) does not look in the area of the CS.  The latent-inhibition group begins looking at the CS, but that orienting habituates.  Notice also the pattern closely matches what one would expect with the operation of two habituation mechanisms (short term and long term).  Within the CS, orienting decreases and then recovers somewhat between presentations.  Overall, between presentations, orienting decreases.  

When not looking at the CS, they are looking at other parts of the screen.  As shown in the earlier heatmaps, the screen is divided into five sections, 1: the circular section around the CS, 2: the upper left, 3: upper right, 4: lower left, and 5: lower right.  The upper left was referred to as the “US Zone” because that was where the US always appeared; somewhere from within the upper left quadrant.  The figure below shows participants’ gaze position on the screen, collapsed across groups because they did not differ in how they distributed their gaze across the rest of the screen in this phase.

On some seconds of some trials there was a significant preference for looking in the US Zone, at least over looking at the bottom right.  These points were likely artifacts of preferring to look to the left of the screen.  Within both the top and bottom sections, the circles (left) are higher than the corresponding triangles (right).  (Orange > blue, red > yellow)

Conditioning

The next figure shows the behavioral keypress response data for the two groups during each second of each trial of conditioning.  All figures in this section show the first 5 seconds of the CS.  After that, the US appears and generates considerable responding.  With the eyes, once the US appears their gaze is focused on the US.

Robust group differences emerged on trial 4 and persisted across the phase.  On trial 9 there were no reliable differences, but there were again on trial 10.  Thus, after 9 trials of conditioning the response levels between the two groups had not  yet converged: Robust evidence of Latent Inhibition was obtained.  Moreover, it is worth noting that such evidence was obtained without the use of a masking task.

Latent Inhibition: First conditioning trial

The first conditioning trial looked more or less like the last pre-exposure trial, with looks at the sensor and other looks distributed across the screen.  However, the Control group was receiving the sensor for the first time…

Control Group: First conditioning trial

…and they directed their attention, as expected, predominately at the sensor.  By the last trial of conditioning, interesting patterns had emerged.

The top image below shows the heatmap for the Latent inhibition group on the last conditioning trial, while the next one following below shows that for the control group.  What each shows is that gaze was directed toward the sensor, and also to the upper left zone of the screen, the US Zone.  Over trials, participants came to expect the US and direct their attention to where it would appear.

Latent Inhibition: Last conditioning trial

Control: Last conditioning trial

Looks at the sensor are summarized below.  The control group clearly looked at the sensor more than did the latent inhibition group.  In the Latent inhibition group, looking at the sensor during the first second of each trial increased after the first trial.  But, overall (see trials 5, 8, and 10) attention to the CS in the latent inhibition did not quite match the level observed in the control group.

When not looking at the CS, participants came to direct their gaze to the area of the screen where the US would appear, as shown by the red lines in the figure below.

There was some trend for the Control group to show this CS-directed goal-tracking response more strongly (see trials 4, 5, 7, & 9: But see trial 8).   These responses are complicated somewhat by the unconditioned tendency in both groups to prefer looking in that area, yet the preference during the CS was considerably more pronounced.

Extinction

The figure below shows the behavioral data for extinction.  At this point, the groups did not differ.  Thus, they converged behaviorally after 10 conditioning trials.  No US was present, so I am showing behavior elicited by the CS over the entire 20-s of the CS.  Both groups showed a timed response.  Their responding increased to the point where the US was expected (indicated by the yellow arrows) and then began to decrease.  Thus, there was a within-trial decrease in responding, likely due to a generalization decrement, and a between-trial decrease likely due to extinction processes.

The eye data at the beginning of extinction did not differ from that at the end of conditioning.  However, the pattern had changed by the end of extinction.

Both the Latent Inhibition Group (below) and the Control group (2nd map below) directed some attention to the CS, but their goal tracking response had extinguished.  When not looking at the CS, they distributed their gaze more evenly across the screen.

 Latent Inhibition: Last extinction trial

Control: Last extinction trial

The figure below summarizes the looks directed at the sensor CS.  It collapses across groups as there were no meaningful differences between groups, and it averages across every two trials to improve clarity.

There was no change in the first second over trials.  The first second still commanded at least a brief 1/4 second glance.  After that, looks to the CS decreased up till the time when the US was expected (indicated by the arrow).  At that point, looking at the CS increased on trials 1-6, during the time the US was expected, but absent.  Once the US did not occur, on trials 7 and 8 there was little change in looking at the CS during that time, and on trials 9 and 10 looking at the CS continued to decrease.  

Extinction produced a small increase in attention to the CS that itself declined as extinction progressed.

The final figure below shows the participants’ gaze over the rest of the screen.  At the start of extinction they looked predominately at the zone where the  US would appear, and that declined over trials as looks at other areas increased.  One group may show more responding than another on some second of some trial in some zone, but overall there are no meaningful group differences here.

Now that you have got this far, the animated version is below.

Byron Nelson presenting at the 2011 congress of the Spanish Society of Comparative Cognition, Sept 21, 2011.

Caveat: These blog entries are generated quickly, and though I do hope to communicate something, their main purpose is to serve as a repository for what I was thinking about particular sets of data that I can refer to when I get around to writing them up. So what you read here, as far as writing goes, is essentially in first and rough draft form. So, while I appreciate feedback, pointing out my grammar errors & whatnot isn’t really useful at this stage.

Richard Sharp Shaver wrote some interesting and fun science fiction for Amazing Stories during and post World War II.  In these stories, he describes ancient people (the Deros and Teros) who lived deep underground and control events on earth through powerful telepathic machines.  However, as he published these stories he claimed that they were, in fact, true occurrences and that only science fiction magazines were willing to publish them.  These stories were presented by Ray Palmer as “The Shaver Mystery.”

It seems, according to Palmer, that many people believed the story and wrote letters corroborating the events Shaver describes.  One of these believers was Charles Marcoux, and it is here that the connection to the Blowing Cave in Cushman, Arkansas begins.  Mr. Marcoux appears to have been told either by the mysterious David L or Mr. Wight that the Blowing Cave was an entrance to the underground world described by Shaver.   This is, of course, interesting to me because at least three generations of my family come from Cushman, and I have myself been in that cave several times.

Entrance to the Blowing Cave in Cushman Arkansas

The cave has a large opening which goes back about 100 yards, then one must begin to crawl through openings in rubble.  Once inside, there are areas where one can stand again, pools of water, and small waterfalls.   According to stories told by my grandfather, it was once possible to walk much farther back into the cave before it was required to enter crawlspaces.  He told stories of dances being held in the cave being lit by the headlights of model T Fords with the floor cracking and the Fords being lost as the floor gave in (hard to tell how much of Grandpa’s stories were fact, fiction, or an exaggerated combination of both).  Although the end of accessible portions of the cave seem to be about a mile, to 1.25 miles deep now, Grandpa said that teams had spent days in the cave in the past without finding the end.  This latter statement is believable as the cave, in my experience, is constantly changing.   Perhaps when Marcoux exposed the entrance to their underground kingdom, the dwellers closed it in.

Once my wife learned of the stories in the links I am sharing below, and knowing my connections to that area she proclaimed, “Well- that explains a lot.” (chuckle)

If someone in interested in the story, a good place to start is  here.

For information regarding Marcoux’s claims and experiences, the N Chronicles is a fun read, as is this excerpt from “The Underground Empire”  “Bury me in Blowing Cavern” provides some synopsis as well.  And this proves to be interesting.  It seems that there is quite an internet culture surrounding the Shaver Mystery and Marcoux.  These beings are supposed to have an influence on almost everything we on the surface do, to protect themselves from discovery, including the Kennedy Assassination.

This page contains a great deal of information, as well as related stories.

And This page deals with it from another angle.

Have fun, and if you are familiar with any of this stuff, or Blowing Cave stories in general I would like to hear about them.

In short, latent inhibition was robustly obtained.  Pre-exposed participants showed an orienting response to the CS which habituated over trials.  During conditioning, the control group (not pre-exposed) showed robust orienting to the CS, which declined over trials.  The latent inhibition group (pre-exposed to the CS) showed a small recovery of orienting, but never really quite to the level being displayed by the control group.

Both groups showed a US expectancy conditioned response of looking into the area of the screen where the US was going to appear, with weak evidence that the pre-exposed group was less likely to exhibit this response.  That response emerged over trials, and decreased during extinction.

Imagine you have a complicated data structure.  You have an object that contains many other objects, and then the first object belongs to a list.  For example, in my Eye Tracker Analysis tool I have an object “AOI” (area of interest) that contains various other data structures such as a string “Name,” an array of integers to determine where it is the screen “origin”, a list of integers for “control points” to control its shape and so forth.  Then, I have a list of these AOI structures to use when defining more than one area of interest.

Once created, I need my program to save these areas of interest to a file so that I can later re-load them without having to manually create them again.   In the old days, I would have had to write a routine that manually exported each element in an AOI to a file, and then do the opposite to read them back in.  It would have been a lot of really tedious and un-interesting coding.

Enter XLM serialization. 

To save my list of AOIs, all that was required was to use the XmlSerializer class, and create one that serializes my AOI datatype.  That was really hard…

XmlSerializer serialfile = .XmlSerializer(typeof(AOI_to_Serialize) );

…where (AOI_to_Serialize) is my name for a class that holds the list of my AOI structures.

Then, open a file.

TextWriter tofile = new StreamWriter(filename);

“serialize” your datastructure to that file.

    serialfile.Serialize(tofile,aoi_to_serialize);

And close the file.

        tofile.Close();

You get your data structure outputted to an XML file.  Then, if you want to read that saved structure back into the program, it’s a process parallel to the one just described, but you’ll call  .DeSerialize instead. 

Tons and tons of code reduced to a few lines of .Net wonder.