The paper analyzes some properties of denoising autoencoders using the problem of misspellings correction as an exemplary task. We evaluate the capacity of the network in its classical feed-forward form. We also propose a modification to the output layer of the net, which we called multi-softmax. Experiments show that the model trained with this output layer outperforms traditional network both in learning time and accuracy. We test the influence of the noise introduced to training data on the learning speed and generalization quality. The proposed approach of evaluating various properties of autoencoders using misspellings correction task serves as an open framework for further experiments, e.g. incorporating other neural network topologies into an autoencoder setting.
Authors
Additional information
- DOI
- Digital Object Identifier link open in new tab 10.1007/978-3-319-67077-5_42
- Category
- Aktywność konferencyjna
- Type
- materiały konferencyjne indeksowane w Web of Science
- Language
- angielski
- Publication year
- 2017
