- author = "Santiago Gonzalez and Risto Miikkulainen",
- title = "Optimizing Loss Functions Through Multi-Variate {Taylor} Polynomial Parameterization",
- booktitle = "Proceedings of the 2021 Genetic and Evolutionary Computation Conference",
- year = "2021",
- editor = "Francisco Chicano and Alberto Tonda and Krzysztof Krawiec and Marde Helbig and Christopher W. Cleghorn and Dennis G. Wilson and Georgios Yannakakis and Luis Paquete and Gabriela Ochoa and Jaume Bacardit and Christian Gagne and Sanaz Mostaghim and Laetitia Jourdan and Oliver Schuetze and Petr Posik and Carlos Segura and Renato Tinos and Carlos Cotta and Malcolm Heywood and Mengjie Zhang and Leonardo Trujillo and Risto Miikkulainen and Bing Xue and Aneta Neumann and Richard Allmendinger and Fuyuki Ishikawa and Inmaculada Medina-Bulo and Frank Neumann and Andrew M. Sutton",
- pages = "305--313",
- address = "internet",
- series = "GECCO '21",
- month = jul # " 10-14",
- organisation = "SIGEVO",
- publisher = "Association for Computing Machinery",
- publisher_address = "New York, NY, USA",
- keywords = "genetic algorithms, genetic programming, ANN, Neural networks, Loss Functions, Metalearning, Evolutionary Strategies",
- isbn13 = "9781450383509",
-
URL = "
https://nn.cs.utexas.edu/downloads/papers/gonzalez.gecco21.pdf",
-
DOI = "
10.1145/3449639.3459277",
- size = "9 pages",
- abstract = "Metalearning of deep neural network (DNN) architectures and hyperparameters has become an increasingly important area of research. Loss functions are a type of metaknowledge that is crucial to effective training of DNNs, however, their potential role in metalearning has not yet been fully explored. Whereas early work focused on genetic programming (GP) on tree representations, this paper proposes continuous CMA-ES optimisation of multivariate Taylor polynomial parameterizations. This approach, TaylorGLO, makes it possible to represent and search useful loss functions more effectively. In MNIST, CIFAR-10, and SVHN benchmark tasks, TaylorGLO finds new loss functions that outperform the standard cross-entropy loss as well as novel loss functions previously discovered through GP, in fewer generations. These functions serve to regularise the learning task by discouraging overfitting to the labels, which is particularly useful in tasks where limited training data is available. The results thus demonstrate that loss function optimization is a productive new avenue for metalearning",
-
notes = "University of Texas at Austin
GECCO-2021 A Recombination of the 30th International Conference on Genetic Algorithms (ICGA) and the 26th Annual Genetic Programming Conference (GP)",
