Data and Analysis Unit: PLYM10

Last update: 2015-04-04 by Andy Wills

Simulation

This section contains a description of, and source code for, the simulation reported in Figure 1 of O'Connell et al. (submitted).

Description

The model architecture and parameters were as specified in Gluck and Myers (1993, Appendix), with the following exceptions: (a) in the current experiment, there was an outcome on every training trial (category A or category B), thus the US-present learning rate was used on all training trials, (b) input-to-hidden connections in the cortical model had a learning rate of 0.1, (c) weights were intialized to a [-3,+3] random uniform distribution.

The model had 24 units representing the stimuli (one for each icon), 12 hidden units, and 2 units representing the outcome (one for each category label). Activation from the outcome units was converted to a response probability using a standard exponential ratio decision mechanism (see e.g. Wills et al., 2000). This decision mechanism's single parameter (k) was arbitrarily set to 3; the conclusions of this simulation are robust to a wide range of values for k.

The model was trained and tested on the same number of trials as were the participants, with the stimuli generated using the same algorithms as the program which ran the experiment. As is standard in applications of the Gluck-Myers model, training was preceded by 500 blank trials (ie. no stimuli and no category label), in order to initialize the network.

Figure 1 of O'Connell et al. (submitted) is an average across 2000 simulated participants, each with a different set of randomly-selected initial weights.The simulation was conducted largely in C, for speed of execution, with non-time-critical collation and graphing performed in R. The source code for this simulation is provided below.

Source code

• plym10sim.zip (10K) - A ZIP archive containing the following files:
  • analysis.R - Run this one - R script that compiles the simulation from C source code, performs 2000 runs of the simulation, calculates the average ouput from these runs, and presents outputs results as a graph in a PDF file ('bigsum.pdf')
  • plotfunc.R - Custom graphing functions used by analysis.R
  • main.c - Top-level source code file for Gluck-Myers simulation.
  • include.h - Sets most model parameters for Gluck-Myers simulation.
  • run.c - Source code for functions to initialize, train, and test Gluck-Myers model.
  • cort.c - Source code for the cortical network of the Gluck-Myers model.
  • hipp.c - Source code for the hippocampal network of the Gluck-Myers model.
  • utilities.c - Source code for a set of utility functions used by other C files in this archive.

In order to run this simulation, you will need to install the following:

• R - a free software environment for statistical computing and graphics.
• gcc - a free compiler. If you did a full install of R, you probably have this already.

This simulation was conducted on a Linux machine. It should work with little or no modification on any UNIX-based system (including Apple's OS X). For MS Windows, you may have to modify the system() calls in analysis.R - happy to help you do this, on request.

Empirical data and analysis

Description

This section publishes such participant data files as are permitted under the ethical and regulatory approvals governing the research reported in O'Connell et al. (submitted). Analysis scripts and stimulus image files are also provided.

Resources

• plym10pat.csv (6K) - Participant-level raw data for patient group and matched control group. Legal/ethical issues mean this raw data cannot be published. It may be possible to provide a confidential copy for research purposes on request - contact me. Column headings are as follows:
  • subj: Participant number (unique within PLYM10)
  • grp: Experiment condition (exp = patient group, ctrl = match controls)
  • datum: Specifies type of data presented in the current row:
    • ph1error: Training phase error score.
    • a12 ... a0: Number of category A elements in the test stimulus.
    • age: Participant's age.
  • value: Dependent on value of 'datum':
    • where datum = 'ph1error': Number incorrect responses in training phase
    • where datum = 'age': Participant's age at time of testing, in years.
    • otherwise: Number of 'category A' responses during test phase.
• plym10umctrl.csv (145K) - Trial-level raw data for unmatched control group. Column headings are as follows:
  • date: Date of testing.
  • subj: Participant number (unique within PLYM10).
  • phase: Experiment phase. 1 = training, 2 = test.
  • trial: Trial number (resets between phases).
  • stim: Phase specific:
    • Phase 1: The category from which the training item was generated.
    • Phase 2: Number of category B elements (0-12).
  • resp: Participant's response.
  • rt: Participant's reaction time (in milliseconds).
• plym10analysis.R (6K) - R script for analyses reported in O'Connell et al. (submitted). This script can be run by the command source('plym10analysis.R'), or can be run interactively. It also produces Figure 2C as a PDF. In addition to the two following files, you will need to have installed the R packages ez and perm. Both are available on CRAN.
  • plym10plot.R (1K) - R function to produce Figure 2C of O'Connell et al. (submitted).
  • bsci.R (2K) - R function for calculation of difference-adjusted between-subject confidence intervals. Implements same calculation as bs.ci() from Baguley (2012). Written by Thom Baguley.
• cam1stim.tbz (11K) - The individual icons making up the stimuli in this experiment. For each participant, 24 icons are randomly selected from the set 1.png to 36.png (37.png to 40.png are not used).

References

• Baguley, T. (2012). Calculating and graphing within-subject confidence intervals for ANOVA. Behavior Research Methods, 44, 158-175.
• Gluck, M.A. and Myers, C.E. (1993). Hippocampal mediation of stimulus representation: A computational theory. Hippocampus, 3,491-516.
• O'Connell, G., Myers, C.E., Hopkins, R.O., McLaren, R.P., Gluck, M.A. and Wills, A.J. (submitted). Amnesic patients show superior generalization in category learning.
• Wills, A.J., Reimers, S., Stewart, N., Suret, M. and McLaren, I.P.L. (2000). Tests of the ratio rule in categorization. Quarterly Journal of Experimental Psychology, 53A,983-1011.