Corrigendum: Gradient-free training of recurrent neural networks using random perturbations
Jesús García Fernández, Sander Keemink, Marcel A. J. van Gerven · Frontiers in Neuroscience · 2024
A Corrigendum on Fern ández, Jes ús García, Sander Keemink, and Marcel van Gerven. In the published article, there was an error in Subfigures 5B and 5C as published.The training and test sequences for this task were identical due to an issue with seed initialization 8 in the dataset generation function. This occurred due to a mistake merging old and new code, 9 which caused an unintentionally re-initialized seed every time a sequence was generated. Thankfully, 10 the impact of this mistake is very small. The updated results show almost no change in overall 11 performance. In fact, the slightly different results for this task now better align with those obtained 12 in the other tasks (the Mackey-Glass time series and the real-world weather prediction task), further 13 strengthening our claims about the robustness and generalization of our proposed approach (ANP).The corrected Subfigures 5B and 5C and its caption "Copying memory data and results. a) At the 15 top, we depict an example of an input with annotations. The sequence length is 20 and the delay 16 period is 10. At the bottom, we show the predictions of a BP-trained model before and after training.