Using Testing to Enhance Learning: A Comparison of Two Hypotheses: (537052012-223)

Michael C. Mozer, Harold Pashler, Michael J. A. HoweeScholarship 2014

Abstract

Students learning facts such as foreign language vocabulary often rely on a self-testing procedure in which they cue themselves with the English word and try to recall the foreign language target, instead of simply memorizing cue-target pairs. The value of this strategy has been empirically verified by a long history of research, yet existing computational models of human learning do not address the enhancing-learning-through-testing phenomenon. Using a simple, well studied model—a feedforward neural net with no hidden units—we propose two different hypotheses for characterizing the phenomenon. Hypothesis 1 is that self-testing generates a target which is used for additional training. Hypothesis 2 is that self-testing produces a more reliable error signal for training.

Introduction

The "testing effect" is a well-documented psychological phenomenon where the act of retrieving information (Test-then-Study, or TS) leads to better long-term retention than simply re-studying the material (Study-Only, or SO). While behavioral evidence for this effect is robust, the underlying computational mechanism remains debated. This paper explores whether testing acts as an additional practice trial or if it fundamentally alters the quality of the learning signal.

Computational Modeling of the Testing Effect

The researchers employed a feedforward neural network without hidden units, trained using the Least Mean Squares (LMS) learning rule. To investigate the mechanism, they modeled neurons as "slow integrators," allowing for a temporal gap between the presentation of a cue and the arrival of a target.

Two primary hypotheses were tested:

  1. Hypothesis 1 (Self-Generated Training): Testing provides an extra training trial where the learner's own guess acts as a target.
  2. Hypothesis 2 (Complete Processing of Cue): Testing allows the network to fully process the cue before the target is presented. In SO conditions, the immediate arrival of the target "contaminates" the internal representation of the cue, leading to a noisier error signal.

Results and Analysis

The model's performance was compared against human behavioral data from Carrier and Pashler (1992). The simulations demonstrated that Hypothesis 2 accurately replicated the performance gap observed in human subjects, whereas Hypothesis 1 failed because reinforcing early, incorrect guesses proved counterproductive.

Comparison of human experimental data and model performance across three experiments.

As shown in the table above, the model's Mean Squared Error (MSE) closely mirrors the error rates found in human trials, confirming that the TS advantage is a robust outcome of the "complete processing" mechanism.

Temporal Dynamics

A key prediction of Hypothesis 2 is that the benefit of testing depends on the timing between the cue and the target. As the "Cue-Target Asynchrony" increases, the network has more time to form a stable representation of the cue, leading to a more precise error signal when the target finally appears.

Mean Squared Error as a function of Cue-Target Asynchrony for Study-Only (SO) and Test-then-Study (TS) conditions.

The model shows that while both conditions benefit from longer processing times, the TS condition maintains a significant advantage, particularly as the delay allows for full integration of the cue.

Discussion

The success of Hypothesis 2 suggests that the testing effect is not merely about "extra practice." Instead, it highlights the importance of signal clarity in error-correction learning. By delaying the target, testing ensures that the resulting error signal—the difference between the predicted and actual target—is based on a fully developed internal state. This finding aligns with theories of Temporal Difference (TD) learning, where the brain's primary goal is to generate accurate predictions of future stimuli as early as possible.

Conclusion

This research bridges the gap between behavioral psychology and computational neuroscience, providing a plausible mechanism for one of the most effective study strategies. Future work may extend these findings to more complex, deep architectures to better understand how these principles apply to structured knowledge domains like language acquisition.

Summary updated

Abstract

Students learning facts such as foreign language vocabulary often rely on a self-testing procedure in which they cue themselves with the English word and try to recall the foreign language target, instead of simply memorizing cue-target pairs. The value of this strategy has been empirically verified by a long history of research, yet existing computational models of human learning do not address the enhancing-learning-through-testing phenomenon. Using a simple, well studied model—a feedforward neural net with no hidden units—we propose two different hypotheses for characterizing the phenomenon. Hypothesis 1 is that self-testing generates a target which is used for additional training. Hypothesis 2 is that selftesting produces a more reliable error signal for training
BibTeX
@inproceedings{mozer2014using,
  author = {Michael C. Mozer and Harold Pashler and Michael J. A. Howe},
  title = {{Using Testing to Enhance Learning: A Comparison of Two Hypotheses: (537052012-223)}},
  booktitle = {eScholarship},
  year = {2014},
}
Figure