Learning and Forgetting Speed Parameter

At the moment, all our learning modules learn at the same speed (instantaneously integrating information into their models). However, as Monty becomes more hierarchical, it will likely be necessary to have adjustable learning speeds where some LMs learn slower and others faster. For instance, we want to be able to quickly learn a representation of the current environment to be able to interact with it (and also quickly forget it again). This is similar to what happens in the hippocampus and is what Monty can currently do (without the forgetting part). However, we will also need more stable models that don't change quickly and are remembered for a long time. This is important since higher-level LMs rely on the output of lower-level LMs. If that output is changing faster than the higher level models can adapt, they will fail to learn useful compositional models.

The idea here is to leverage the existing constrained object models and add a learning speed parameter into it. The concrete way in which this parameter could work would need to be outlined in an RFC. Ideally, it should also incorporate forgetting.

Help Us Make This Page Better

All our docs are open-source. If something is wrong or unclear, submit a PR to fix it!

Make a Contribution

Learn how to contribute to our docs