Another SRN issue, tracking accuracy

[Stephen Read]

Hi Michael,

Following some questions in class today, we discovered that  if you train the SRN on the First.tsv data set and then look at the results on the TrainEpochPlot it looks like the network learns perfectly, even without a hidden layer, but also with a hidden layer, with SSE being 0 and accuracy perfect. The  READ.ME says it should not be able to learn a first order dependency in either case.  However, if you also do a testing run, as shown in the attached QT movie you will see that it actually does not learn the second order dependencies.  If you look at the network and the activations you will see that both times B appears, it activates A and C equally, instead of activating A one time and C the other.  (when I say A and B in the video, I should have said A and C). Also, if you click on Test Trial, just to the right of Test Epoch Plot button you will get a table of results which shows two errors.  So in doing the homework, you evidently can’t trust the Train Epoch Plot graph, but should instead do a test run as I show in the video. Also, the Test Epoch Plot button also shows that the network is not learning.

   I seem to remember that Randy mentioned somewhere that he had changed code for determining whether a prediction was accurate, by using the criteria of .5 or greater activation.  I don’t know if that is what is responsible for this, but regardless it doesn’t look the code is calculating accuracy properly.

Best, Steve

https://www.dropbox.com/scl/fi/3o2g0186sf07xfdpwher5/Training-Testing_SRN.mov?rlkey=lgms32qb7yrvlf7crytrj6g0h&st=yed4tckl&dl=0

Stephen J. Read

Mendel B. Silberberg Professor of Social Psychology

Department of Psychology

University of Southern California

Los Angeles, CA 90089-1061

Website: www.stephenjread.com

 

 

 

[Michael J Frank]

 Hi Steve, cc'ing Krishn has he was TA for me last year and may remember some additional things. And obviously Randy can chime in. 

But two points

1. it is possible the threshold for counting a trial as accurate is too liberal - you can simply go to the srn.go code and change

 sse, avgsse := out.MSE(0.5)

to

sse, avgsse := out.MSE(0.25)

or so and then it will require the correct unit to be within .25 accuracy. That should make it clear the network isn't learning in that case without a hidden or SRN layer.

2. ANother thing we noticed last year is that the network can solve a first order sequence without a context layer - not as reliably but still it could which was weird. It turned out that this was because in the newer versions even though there is no context, the Ge values have some hysteresis from one trial to the next.  So in the very first cycle of trial 2, regardless of the input, the Ge values in the neurons are biased to be higher for those that had highest Ge at end of last trial. This is presumably related to GTau which probably doesn't get reset even if activations do across trials.  As such the network can capitalize on this trace bias to have some memory. 

I don't remember if we fixed that or just explained it to the students as to why there is still some success (it effectively acts like a context layer in its Ge memory). So this is a heads up to look for, and maybe could be addressed by changing those Ge params.

[Stephen Read]

Thanks Michael, that worked and thanks for the info in number 2.  

I suspect that the problem I was having with the incompatibility of the srn with the 1.25 series compiler might have been due to the accompanying go. mod specifying an older version of some of the required libraries. Do you know if that is true?

Best Steve

Stephen J. Read

Mendel B. Silberberg Professor of Social Psychology

Department of Psychology

University of Southern California

Los Angeles, CA 90089-1061

Website: www.stephenjread.com

 

 

 

From: Michael J Frank <Michae...@brown.edu>
Date: Thursday, November 6, 2025 at 5:10 AM
To: Stephen Read <re...@usc.edu>
Cc: Randall O'Reilly <rcore...@gmail.com>, emergent-users <emergen...@googlegroups.com>
Subject: Re: Another SRN issue, tracking accuracy

This Message Is From an External Sender

USC detected that this message came from outside of USC.

Report Suspicious

 

[Michael J Frank]

Krishn --- does that ring a bell (older libraries for go.mod for last year's midterm)?

 Steve - are you saying that this is fixed if you just use a newer go.mod?

[Stephen Read]

No I haven’t tried it. But I do know at least one of the students got it to work with 1.25. And several got a message about one of required packages being incompatible with 1.25. 

Get Outlook for iOS

---

From: emergen...@googlegroups.com <emergen...@googlegroups.com> on behalf of Michael J Frank <Michae...@brown.edu>
Sent: Friday, November 7, 2025 7:14:58 AM
To: Stephen Read <re...@usc.edu>; Krishn Bera <krish...@brown.edu>
Cc: Randall O'Reilly <rcore...@gmail.com>; emergent-users <emergen...@googlegroups.com>

Subject: Re: Another SRN issue, tracking accuracy

 

Krishn --- does that ring a bell (older libraries for go. mod for last year's midterm)? Steve - are you saying that this is fixed if you just use a newer go. mod? On Thu, Nov 6, 2025 at 1: 42 PM Stephen Read <read@ usc. edu> wrote: Thanks

--
You received this message because you are subscribed to the Google Groups "emergent-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to emergent-user...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/emergent-users/CAGYaZ%2BdC4OeTRZWkQTp9h0CLQr2T%2B4q5a2h5%2BYUamAQnmN%3Dx8g%40mail.gmail.com.