On the Need for a Neural Abstract Machine

Diego Sona, Alessandro Sperduti

Abstract

The complexity of learning tasks and their variety, as well as the number of different neural networks models for sequence learning is quite high. Moreover, in addition to architectural details and training algorithms peculiarities, there are other relevant factors which add complexity to the management of a neural network for the adaptive processing of sequences. For example, training heuristics, such as adaptive learning rates, regularization, and pruning, are very important, as well as insertion of a priori domain knowledge. All these issues must be considered and matched with the complexity of the application domain at hand. This means that the successful application of a neural network to a real world domain has to answer to several questions on the type of architecture, training algorithms, training heuristics, and knowledge insertion, according to the problem complexity.
BibTeX
@inproceedings{sona2001need,
  author = {Diego Sona and Alessandro Sperduti},
  title = {{On the Need for a Neural Abstract Machine}},
  booktitle = {Sequence Learning},
  year = {2001},
  doi = {10.1007/3-540-44565-x\_7},
}

Loading related content...

Figure