Replacing training on natural stimuli with pre-training on retinal waves accelerates learning for LSTM and GRU architectures.

<b>A.</b> Learning trajectories of three LSTM models on the simulated corridor training (left) and validation (right) datasets. Shown is the average MSE across five-fold cross-validation and the corresponding standard deviation. Note that the performance assessment for the model pre-trained for one epoch on retinal waves (<i>DirectedProp one epoch</i> in pink) starts at epoch two, ensuring direct comparability to the <i>NotPretrained</i> model as this model does not have an advantage in training time. Training continues beyond the epochs shown, with each model stopping individually once overfitting is detected. <b>B.</b> Same as panel A, but for GRU architectures. (PDF)