Scientists crack the code as to how modular design can evolve. A potential AI milestone?
Our computers, our cars, our societies, our bodies, our automobile traffic, our network traffic, our brains and even our own solar system – all modular by design. Smaller mechanisms to complete a larger function. Where did it come from? And the bigger question – how did it happen?
Jeff Clune , Hod Lipson , and Jean-Baptiste Mouret , with support from the National Science Foundation and the Agence Nationale de la Recherche (ANR), believe to have these answers. Published today by the Proceedings of the Royal Society, their results from computational evolution experiments have been released that supports the hypothesis that it is due to the price of resource consumption by the network as a whole. Once there is a “cost” for communication applied to the algorithm, the system immediately and abruptly breaks itself down into a more modular design to run more efficiently.
When it evolves, or designs itself this way, you can think of it as networks becoming less of a 2-dimensional web, and more of a complex 3D lattice, the sort of system that can most commonly be associated with neural network.
Finding support in how to evolve modularity can be seen as a possible holy grail to many fields, but what it can do for AI seems to be the most exciting. By being able to naturally evolve modular design into existing systems, the functionality and complexity of those systems increases significantly.
I asked Jeff Clune and his thoughts on how this impacts Artificial Intelligence, and he had this to say
The field of evolving AI has produced a lot of impressive results, often beating all other machine learning algorithms and outperforming designs made by hand by human engineers. However, especially in the field of evolving neural networks (NNs), the evolved NNs pale in comparison to the sophistication and abilities of natural NNs (i.e. animal brains). One of the main limitations is that natural NNs are structurally organized in a modular design. However, computationally evolved NNs rarely, if ever, spontaneously evolve to have a modular design. A leading view in our field is that the lack of modularity serves as a “complexity ceiling” that limits the level of intelligence that can evolve. With this discovery, we can break through that complexity ceiling and greatly improve the intelligence of evolved computational neural networks.
References – J. Clune, J.-B. Mouret, H. Lipson. The evolutionary origins of modularity. Proceedings of the Royal Society B: Biological Sciences, 2013; 280 (1755): 20122863 DOI: 10.1098/rspb.2012.2863