Dataset Information

Neural modularity helps organisms evolve to learn new skills without forgetting old skills.

ABSTRACT: A long-standing goal in artificial intelligence is creating agents that can learn a variety of different skills for different problems. In the artificial intelligence subfield of neural networks, a barrier to that goal is that when agents learn a new skill they typically do so by losing previously acquired skills, a problem called catastrophic forgetting. That occurs because, to learn the new task, neural learning algorithms change connections that encode previously acquired skills. How networks are organized critically affects their learning dynamics. In this paper, we test whether catastrophic forgetting can be reduced by evolving modular neural networks. Modularity intuitively should reduce learning interference between tasks by separating functionality into physically distinct modules in which learning can be selectively turned on or off. Modularity can further improve learning by having a reinforcement learning module separate from sensory processing modules, allowing learning to happen only in response to a positive or negative reward. In this paper, learning takes place via neuromodulation, which allows agents to selectively change the rate of learning for each neural connection based on environmental stimuli (e.g. to alter learning in specific locations based on the task at hand). To produce modularity, we evolve neural networks with a cost for neural connections. We show that this connection cost technique causes modularity, confirming a previous result, and that such sparsely connected, modular networks have higher overall performance because they learn new skills faster while retaining old skills more and because they have a separate reinforcement learning module. Our results suggest (1) that encouraging modularity in neural networks may help us overcome the long-standing barrier of networks that cannot learn new skills without forgetting old ones, and (2) that one benefit of the modularity ubiquitous in the brains of natural animals might be to alleviate the problem of catastrophic forgetting.

SUBMITTER: Ellefsen KO

PROVIDER: S-EPMC4383335 | biostudies-literature | 2015 Apr

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Neural modularity helps organisms evolve to learn new skills without forgetting old skills.

Ellefsen Kai Olav KO Mouret Jean-Baptiste JB Clune Jeff J

PLoS computational biology 20150402 4

A long-standing goal in artificial intelligence is creating agents that can learn a variety of different skills for different problems. In the artificial intelligence subfield of neural networks, a barrier to that goal is that when agents learn a new skill they typically do so by losing previously acquired skills, a problem called catastrophic forgetting. That occurs because, to learn the new task, neural learning algorithms change connections that encode previously acquired skills. How networks ...[more]

PMID: 25837826

Similar Datasets

Project description:BackgroundIn a recent work I transformed a complex and integrated text like the International Code of Zoological Nomenclature into a network of interconnected parts of text. This new approach allowed understanding that a continuous body of text cannot accurately reflect the true structure of the Code, and provided a scientific methodology to identify a priori parts that could be affected by future revisions. In this next step, I investigate further the structure of the Code, seeking to use the network in order to identify the various conceptual communities grouping the various articles and other text items of the Code.MethodsUsing the first version of the network of the Code, I perform a comprehensive modularity analysis in two rounds: the first round aims to identify the fewest and largest communities or modules for the entire network, whereas the second round identifies the sub-modules within each larger module. The potential conflicts between the current structure of the Code and the module composition are evaluated with a parcellation analysis.ResultsThe optimal modularity search identified 10 different modules in the entire network of varying size (ranging from 75 to 200 nodes). Each module can be further divided into smaller modules, that all-together allow describing the 65 conceptual groups of text items in the Code. Parcellation analysis revealed that two-thirds of the current chapters of the Code are in excellent or good accordance with the recovered conceptual modules, whereas the current composition of six chapters is in serious conflict with the conceptual structure of the Code.DiscussionJudging only the composition and not the order of appearance of the Articles in the Chapters of the Code, I show that in many cases the current structure of the Code is found to correspond quite well to the concepts presented therein. The most important conflict is found on the provisions related to the various groups of names governed by the Code: family-, genus-, and species-group names. Currently, these provisions are spread out in different Articles in different Chapters, along the entire length of the Code. The modularity analysis suggests that re-organizing the Code in chapters that will deal with all aspects related to a given group (e.g., chapters including information on name formation, availability, typification, and validity for a given group), could potentially improve reader experience and, consequently, the applicability of the Code.

Dataset Information

Neural modularity helps organisms evolve to learn new skills without forgetting old skills.

Publications

Neural modularity helps organisms evolve to learn new skills without forgetting old skills.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets