Publication: Evolutionary design of deep neural networks
Loading...
Identifiers
Publication date
2018-06
Defense date
2018-11-05
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
For three decades, neuroevolution has applied evolutionary computation to the optimization of
the topology of artificial neural networks, with most works focusing on very simple architectures.
However, times have changed, and nowadays convolutional neural networks are the industry and
academia standard for solving a variety of problems, many of which remained unsolved before the
discovery of this kind of networks.
Convolutional neural networks involve complex topologies, and the manual design of these
topologies for solving a problem at hand is expensive and inefficient. In this thesis, our aim is to
use neuroevolution in order to evolve the architecture of convolutional neural networks.
To do so, we have decided to try two different techniques: genetic algorithms and grammatical
evolution. We have implemented a niching scheme for preserving the genetic diversity, in order
to ease the construction of ensembles of neural networks. These techniques have been validated
against the MNIST database for handwritten digit recognition, achieving a test error rate of 0.28%,
and the OPPORTUNITY data set for human activity recognition, attaining an F1 score of 0.9275.
Both results have proven very competitive when compared with the state of the art. Also, in all
cases, ensembles have proven to perform better than individual models.
Later, the topologies learned for MNIST were tested on EMNIST, a database recently introduced
in 2017, which includes more samples and a set of letters for character recognition. Results have
shown that the topologies optimized for MNIST perform well on EMNIST, proving that architectures
can be reused across domains with similar characteristics.
In summary, neuroevolution is an effective approach for automatically designing topologies for
convolutional neural networks. However, it still remains as an unexplored field due to hardware
limitations. Current advances, however, should constitute the fuel that empowers the emergence of
this field, and further research should start as of today.
Description
Mención Internacional en el título de doctor
Keywords
Topology design, Neuroevolution, Evolutionary computation, Convolutional neural networks, Genetic algorithms, Grammatical evolution, Representation learning, Deep learning, Machine learning, Artificial intelligence