CECN1 Act Priming

From Computational Cognitive Neuroscience Wiki

Jump to: navigation, search

Activation Based (Short Term) Priming

Back to CECN1 Projects

Project Documentation

(note: this is a literal copy from the simulation documentation -- it contains links that will not work within the wiki)


  • GENERAL USAGE NOTE: To start, it is usually a good idea to do Object/Edit Dialog in the menu just above this text, which will open this documentation in a separate window that you can more easily come back to. Alternatively, you can just always return to this document by clicking on the ProjectDocs tab at the top of the middle panel.

This short project is essentially just the same simulation as the weight priming one we did earlier in the chapter (wt_priming.proj), so we will assume you are already familiar with the network and environment. We begin by loading a pre-trained network, although you are welcome to train one from scratch by doing Train: Init and Run.

  • Press the Load Weights button at the bottom of the master .PanelTab.ControlPanel, which loads in the weights file called act_priming_trained.wts.gz

Let's test this network (using a somewhat different setup than last time) to obtain a baseline measure of performance, where we do not expect any activation-based priming (because we clear out the activations after each trial, due to the decay.event parameter being set to 1 initially).

This test has the 'a' and 'b' responses to a given input being presented one after another, which allows us to determine the immediate impact of seeing the 'a' case on the response to 'b'. When the 'a' case is presented, we clamp the output response pattern to the 'a' value. When the 'b' case is then presented, only the input pattern is presented, so the 'b' output response just serves as a comparison pattern for the actual output produced by the network given the input.

  • Press Test: Init, Step in the master .PanelTab.ControlPanel to present the first ('0_a') event to the network. You should see both the input and output pattern clamped
  • Now press Step several more times while keeping your eye on the Output layer to watch the responses. You can see the closest event stats in the bottom of the two grid log views shown above the network -- as you can see, each Step presents one event in the following order: 0_a,0_b,1_a,1_b,2_a,2_b, etc.
  • Press the Test: Run button to finish out the full test epoch of 26 events, at which point you'll get a new row of data in the top grid view, showing the summary for the epoch.

As with the wt_priming model, the key summary value is sum_name_err, which reflects the number of times the network produced the 'a' output for the 'b' inputs (because the 'a' trials have their output patterns clamped, they always have a name_err of 0, so this measure is only for the 'b' trials which have the output unclamped). The number you see here reflects the baseline value, against which we will compare the activation-based priming value.

As noted above, the decay.event parameter in the master .PanelTab.ControlPanel controls the extent to which the activations are decayed (reset) after each event is processed. Let's observe the effects of keeping the activations completely intact from one trial to the next.

  • Change the decay.event parameter to 0 instead of 1. Then do Test: Step for a few trials, and observe how the residual activations influence the processing of the output patterns in the network. Do a Test: Run to finish out the epoch of testing.

You should notice an increased tendency for the network to respond with 'a' in the 'b' trials (also see figure 9.15 in text).

Question 9.8 (a) Report the number of times the network responded 'a' instead of 'b' for the 'b' test trials, relative to the baseline that you observed above with decay.event = 1. (b) Explain why decay.event has this effect on performance.

In summary, this simulation replicates our earlier weight-based priming results using residual activations instead of weight changes, and illustrates a simple form of activation-based memory that our model of the cortex naturally exhibits.

  • When you are done with this simulation, you can either close this project in preparation for loading the next project, or you can quit completely from the simulator.
Retrieved from "http://grey.colorado.edu/CompCogNeuro/index.php/CECN1_Act_Priming"
Personal tools