- Title: Exploring an Evolved Neuron for Human Activity Recognition
- Subtitle: Empirical Study on Performance and Generalization
Recurrent neural nets are one of the key technologies for machine learning and AI. Chau Tran was given a novel neuron evolved by the ADATE automatic programming system using a standard LSTM neuron as the initial individual. The result is a very different neuron that is much better for activity recognition and also for most of the other seven datasets that Chau Tran used for evaluation.
She manually evolved the neuron further and managed to simplify the evolved neuron as well as produce slightly better variants. The ease with which this manual evolution could be done shows that neurons are surprisingly easy to evolve. To know that evolvability is not only possible but rather simple provides a promising path towards Artificial General Intelligence (AGI).