2

I started researching the subject of self-replication in neural networks, and unexpectedly I saw that there is not much research on this subject. I should mention I am new in the field of NNs.

This idea seems to be very appealing, but now I am having problems coming up with an actual use case. The paper Neural Network Quine from 2018 seems to be one of the main ones addressing this topic.

So, what are the use-cases of self-replicating neural networks? Why isn't this subject more thoroughly researched?

nbro
  • 39,006
  • 12
  • 98
  • 176
Ezeeroc
  • 21
  • 2
  • You asked a very similar question [here](https://ai.stackexchange.com/q/24542/2444). In any case, in the abstract of that paper that you cite, the authors write "_We suggest that a self-replication mechanism for artificial intelligence is useful because it introduces the possibility of continual improvement through natural selection._". Moreover, what you are quoting from Reddit, i.e. "it's really just take machine learning model and have it predict its own weights", it's exactly what's written in the abstract of the same paper you're citing. Have you read that paper? – nbro Nov 15 '20 at 12:21
  • I asked again trying to be more specific, as suggested in the other question. Of course I read it. I just wanted to start a discussion on the subject and hear what other (more qualified) people think about it, but it seems I dont know how to ask a question around here to make people respond – Ezeeroc Nov 15 '20 at 18:17
  • 1
    It's also possible that nobody here is familiar with or interested in similar work. That paper actually contains a section "Motivation" where the authors try to motivate their work. I suspect that people still don't see many use cases of these systems, and that's why there may not be much research work (I actually don't know, because I have not researched anything on the topic). It's possible that this type of neural network, which refers to itself, in some way, may turn out to be useful. Later, maybe I will read something about the topic, and maybe I will attempt to provide an answer. – nbro Nov 15 '20 at 18:26
  • Thanks for the answer! Would be helpful if you find the time – Ezeeroc Nov 15 '20 at 19:15
  • This paper [Self-Replication in Neural Networks](https://www.mitpressjournals.org/doi/pdf/10.1162/isal_a_00197) by Thomas Gabor et al. could also be useful to answer your question (I didn't read it yet). – nbro Nov 20 '20 at 13:37
  • 1
    I have read it. It is about tuning NNs to accept other NNs as inputs with the purpose of having themselves as input and thus self-replication. To make this possibile they use several `reductions`. It doesnt really say much about the end goal, or purpose. It is more for research purposes – Ezeeroc Nov 21 '20 at 07:24
  • I'm pasting here the link to the Reddit post that you mentioned: https://old.reddit.com/r/MachineLearning/comments/js4bkz/d_why_arent_self_replicating_neural_networks_more/. – nbro Oct 29 '21 at 10:03
  • @Ezeeroc Have you reached any conclusion which you could share? – Rexcirus Sep 30 '22 at 15:02
  • @Rexcirus only more abstract ones and nothing very specific – Ezeeroc Oct 07 '22 at 05:10

0 Answers0