City Research Online

Abstract

The authors regret to inform the reader that we discovered a programming error that changes some aspects of the results reported in our original study. The amended results are very similar to those originally reported, and our central conclusions are unaffected. When corrected, a preference for ABC units over BC:D units emerges only for a forgetting rate of at least 0.6 (rather than 0.4 in the original report). The results reported in the Supplementary Information as well as the overall conclusions are unaffected. The corrected code is available at http://dx.doi.org/10.25383/city.21294735. In Endress and Johnson (2021), we reported simulations with a neural network of a number of Statistical Learning tasks. We evaluated the network performance by comparing its “familiarity” with different types of test items. We calculated the familiarity of the network with a test item by recording the total network activation, either in the entire network (for the results reported in the main text) or just in the neurons coding for the test items (for the results reported in Supplementary Information D). We compared the network's familiarity with the test items in two ways. For each comparison of test items, we calculated normalized difference scores: [Formula presented] We then (1) compared the difference scores to the chance level of zero using a signed rank test (across simulated participants) and, in analogy to analyses in developmental populations, (2) compared the proportion of positive difference scores to the chance level of 50% using a binomial test; with 100 simulations per parameter set, the chance level is exceeded when at least 61% of the simulations show positive difference scores. (Below, we call difference scores “significant” if they differ from the chance level of zero in a signed rank test. We call the proportion of positive difference scores significant when the proportion of positive (or negative) difference scores differs from the chance level of 50% in a binomial test.) We found that the network reproduced many Statistical Learning results for intermediate forgetting rates, but not for very low forgetting rates or very high forgetting rates. We discovered a programming error that affects the results reported in the main text, while the results reported in Supplementary Information D are correct as reported. We now reran the simulations with the amended coded (available at http://dx.doi.org/10.25383/city.21294735) as well as with the old code (but using current versions of the R libraries required for our simulations). The amended results are very similar to those originally reported, and our central conclusions are unaffected. The main difference to the original results concerns the forgetting rate at which a preference for ABC units over BC:D part-units emerges; these units correspond to words and part-words in linguistic Statistical Learning studies. In the amended simulations, this preference emerges only at forgetting rates of at least 0.6 (rather than 0.4 in the original report). Further, for a forgetting rate of 0.4, a preference for BC:D part-units emerges (see Fig. 1). As in our original report, BC:D part-units are thus harder to reject, but can be rejected with suitable forgetting rates, though the rates need to be slightly higher than in the original report.1 Except for some numerical differences for forgetting rates where learning was unreliable in the original report (i.e., where our evaluation measures above disagreed), the main results as well as the conclusions remain unaffected. In the appendices below, we provide a detailed comparison between the amended and the old results. Specifically, we provide updated versions of Figs. 3, 4, and 5 as well as updated Tables C1 and C2. We also list all changes in the significance pattern. As mentioned above, except for the preference for ABC units over BC:D part-units, these changes occurred exclusively in cases where learning was unreliable in both the original and the amended simulations, in general in cases where the significance pattern was inconsistent between the continuous (signed-rank) and the count-based tests in both the original and the amended simulations. As a result, these changes do not affect our central conclusion that a Hebbian learning model can account for a variety of Statistical Learning results at intermediate forgetting rates. The authors would like to apologise for any inconvenience caused. ADE accepts responsibility for this mistake.

Publication Type:	Article
Additional Information:	© 2023. This manuscript version is made available under the CC-BY-NC-ND 4.0 license https://creativecommons.org/licenses/by-nc-nd/4.0/
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Health & Psychological Sciences School of Health & Psychological Sciences > Psychology
SWORD Depositor:	Symplectic Administrator

Export

Downloads