2024-05-16 Meeting Notes

To discuss

  • Parameter identifiability and trade-offs (more details linked here)
    • The parameter space geometry is complicated
    • Best-fitting parameters differ from the paper reported values
    • I started from 100 different values
    • Possible overfitting?
  • Argument is about best-fitting parameters. But question is more about “reasonable parameters” (related to overfitting)
    • how to determine what range of parameter values are consistent with the data?
  • What is “too slow?”. Comparison with estimates from Popov & Reder (2020) & Ma et al (In press)
    • (!) just model it trial by trial allowing resource to recover however much they can in the time between trials
  • Caputring the primacy effect is to “blame”. Model with multiple contributions to primacy?
    • In experiment 2, after 6 seconds gap there is no difference in SP4-6 between known and random chunks, which the model can only account for width full recovery.
  • Should we model the first chunk?
    • If so, how?
  • How much modeling is enough
    • I need to be convinced that most reasonable versions of the model require very slow rates
  • Conclusions
    • I finally understand the argument for freeing capacity. But now the argument seems one-sided. We do formal modeling to reject the resource account, and based on that we conclude that the other capacity account is more plausible without doing any modeling for it?

Models I’ve tried

  • Model 1a: same as in the paper (with many different starting values)
  • Model 1b: same as 1a but also accounting for the first chunk
  • Model 1c: putting prior on parameters. Particularly on the rate parameter over a grid
  • Model 2: including the encoding time for the recovery
  • Model 3: strength = sqrt(resource_cost)

Conclusions so far: - Despite the complicated geometry and parameter trade-offs, recovery rates do need to be slow with this model group

Modeling left to do

  • Experiment 3
  • Apply model from Ma et al (In press) just replacing LF and HF with random and known chunks
    • analytical likelihood, should be fast, maybe even bayesian
    • can be applied on a trial-by-trial basis
    • (!) just model it trial by trial allowing resource to recover however much they can in the time between trials

Next steps

  • Bootstrap participants and trials to and fit the model too get distribution

    • if possible bayesian

  • Fit the Ma et al (In press) model to the trail level data with actual inter-trail intervals

  • What does the model predict - interaction for introduction

  • Framing - we are testing this model as a potential new comprehensive explanation. Do not embrace alternative conclusion just because (if) this models fails to pass

Back to top