Palestrante
Descrição
Artificial neural networks (ANN) and Complex networks are gaining attention nowadays due to advances in artificial intelligence, computer hardware, and big data phenomena. Recent deep ANN architectures have been achieving outstanding performance on various problems, which led to the proposal of many models. However, due to the lack of knowledge of its internal functioning, these techniques are used as a black box approach. For instance, it is possible to produce images that are completely unrecognizable by humans, but which deep convolutional networks strongly believe to be recognizable objects.(1)New studies are then needed regarding the functioning of these networks, contributing to the development of more robust and interpretable models. An interesting alternative is to analyze ANN as complex systems, such as a complex network. It was shown(2) that an ANN with a small-world topology performs better than traditional fully-connected networks for the diagnosis of diabetes. Moreover, complex network measures show some correlation with the function of the deep belief ANN neurons.(3) In this context, we propose a study to improve the understanding of ANN through complex networks and we hypothesize that the system topology plays a crucial role in its overall functioning. To achieve that, the ANN elements are modeled as a complex network considering the active interaction between its entities (neurons). The topological properties of the system are analyzed with the goal of establishing correlations between its structural organization and functioning. Our first analysis considers deep feedforward ANN as directed graphs, where connections point towards the propagation of information (forward) from the input to the output layer. We then analyze the correlation between two topological measures (input and output strength) of the hidden layers with the network performance (accuracy rate). Results using 1000 ANN with different random initial weights trained in two different datasets indicates a strong correlation between the network accuracy and the average strength of hidden neurons. This indicates the existence of topological elements responsible to improve the network performance. For instance, in our case networks with higher accuracy have hidden neurons with negative average input and output strength, while networks with low accuracy show the opposite. We then believe that there are other network properties related to its internal functioning, such as other centrality and path-based measures. The continuation of this research will then focus on the elaboration of new methods to improve ANN architectures according to our findings regarding its topology properties.
Referências
1 NGUYEN, A.; YOSINSKI, J.; CLUNE, J. Deep neural networks are easily fooled: high confidence pre-dictions for unrecognizable images. In: IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITIONM cvpr, 15, 215.Proceedings... Boston, Massachussetts, 2015.p. 427–436.
2 ERKAYMAZ, O.; OZER, M. Impact of small-world network topology on the conventional artificial neural network for the diagnosis of diabetes Chaos, Solitons & Fractals, v.83, p.178–185, 2016. doi: 10.1016/j.chaos.2015.11.029.
3 TESTOLIN, A.; PICCOLINI, M.;SUWELS, S. Deep learning systems as complex networks.2018. Disponivel em:arXiv preprintarXiv:1809.10941. Acesso em: 19.06.2019.
Subárea | Sistemas Complexos |
---|---|
Apresentação do trabalho acadêmico para o público geral | Não |