We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

nlin.AO

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Nonlinear Sciences > Adaptation and Self-Organizing Systems

Title: On the potential for open-endedness in neural networks

Abstract: Natural evolution gives the impression of leading to an open-ended process of increasing diversity and complexity. If our goal is to produce such open-endedness artificially, this suggests an approach driven by evolutionary metaphor. On the other hand, techniques from machine learning and artificial intelligence are often considered too narrow to provide the sort of exploratory dynamics associated with evolution. In this paper, we hope to bridge that gap by reviewing common barriers to open-endedness in the evolution-inspired approach and how they are dealt with in the evolutionary case - collapse of diversity, saturation of complexity, and failure to form new kinds of individuality. We then show how these problems map onto similar issues in the machine learning approach, and discuss how the same insights and solutions which alleviated those barriers in evolutionary approaches can be ported over. At the same time, the form these issues take in the machine learning formulation suggests new ways to analyze and resolve barriers to open-endedness. Ultimately, we hope to inspire researchers to be able to interchangeably use evolutionary and gradient-descent-based machine learning methods to approach the design and creation of open-ended systems.
Comments: 17 pages, 6 figures; Final author version accepted for publication in Artificial Life
Subjects: Adaptation and Self-Organizing Systems (nlin.AO); Artificial Intelligence (cs.AI); Populations and Evolution (q-bio.PE)
Cite as: arXiv:1812.04907 [nlin.AO]
  (or arXiv:1812.04907v1 [nlin.AO] for this version)

Submission history

From: Nicholas Guttenberg [view email]
[v1] Wed, 12 Dec 2018 11:39:38 GMT (427kb,D)

Link back to: arXiv, form interface, contact.