4. CONCLUSION

The thought experiment presented in this paper is intended to show the theoretical possibility of certain kinds of `self-referential' weight matrices. The network I have described can, besides learning to solve problems posed by the environment, also use its own weights as input data and can (in principle) learn new algorithms for modifying its weights in response to the environmental input and evaluations. This effectively embeds a chain of `meta-networks' and `meta-meta-...-networks' into the network itself.

Back to Metalearning page
Back to Recurrent Neural Networks page