Gating regulates the information retained, forgotten, or incorporated into a recurrent neural-network state. In molecular generation, it helps preserve long-range dependencies needed for valid SMILES, including ring closures, parentheses, and global valence consistency. section 10.4.2
A gated recurrent unit uses two filters:
- The reset gate controls how much of the previous hidden state contributes to the candidate state. section 10.4.2
- The update gate controls how much of the candidate state replaces the previous hidden state. section 10.4.2
- This selective information flow improves long-range dependency handling compared with a simple RNN, but increases computation and memory requirements. section 10.4.2
Molecular sequence memory
When processing benzene as "c1ccccc1" or a longer molecule such as "CC(C)CC(=O)NC1=CC=C(C=C1)O", gating helps retain information about openings that must later be closed correctly. section 10.4.2





