Autores
Sebastian Narvaez, Angel Garcia, Raul Ernesto Gutierrez
Fecha de publicación
2017-06-19
Conferencia
International Work-Conference on the Interplay Between Natural and Artificial Computation, pages 45-52
Editor
Springer, Cham
Abstract
Communicating with machines in the same way we do with other people has been a long-time goal in computer science. One of its many advantages would be the ability to give instructions to our computers without the need of learning how to use specific software or programming languages. Since we’re dealing with human language, it would make sense to use a model of the human brain to build a system with such capabilities. In this work, the Hierarchical Temporal Memory algorithms are explored and evaluated as a biologically inspired tool capable of working with natural language. It’s proposed that task execution can be achieved by training the algorithms to map certain sentences with keywords that correspond to the tasks. Different encoders are tested, that translate words into a proper representation for the algorithms. The configuration of algorithms and encoders with the highest success rate is able to correctly map up to 90% of the sentences from a custom training set. The behaviour of the success rates does not vary greatly between different subsets of the training set, suggesting that the learning system is able to find patterns and make inferences about missing data.