Figure 4
Effect of network size on recovering time parameters. The number of hidden units in the trained model N was varied to test the effect on the loss landscape. The target model that generated the data of these simulations was equipped with rate constants (\(\alpha _s=0.34\), \(\alpha _r=0.68\)) as indicated with the green circle, and 10 hidden units. Networks with 5, 10, 30 and 100 hidden units were trained on the generated data. Both a grid search over fixed rate constants was performed and networks with adaptive rate constants were trained with their final learned rate constants indicated as blue crosses (20 repetitions). The dashed lines indicate the approximation where either \({\hat{\alpha }}_s= 1\) or \({\hat{\alpha }}_r= 1\). The intersection of both dashed lines indicates the Elman solution (blue circle). A larger network size did not decrease the ability of the network to recover the underlying target rate constants.
