Projektdetaljer
Beskrivelse
The goal of GROW is to propose a fundamental new paradigm of machines that are not trained to per- form a certain task but instead can evolve and continually learn from experiences and apply previously learned knowledge to novel situations. Artificial Intelligence (AI) methods are becoming part of our daily lives, in face recognition, speech recognition in mobile phones or automatic translation. Comput- ers can now outperform humans in many domains such as Chess, Go, or even poker. However, these systems still pale in comparison to even simple biological intelligence, which can learn and adapt to unforeseen experiences. Current machine learning systems can only deal with situations they have been trained for in advance; they are unable to adapt during execution to unexpected events that were not anticipated by the designer. In nature, both evolution and neural development (the process that me- diates the formation of new synapses in the brain based on sensory experience) have been shown to be fundamental to learning and the construction of biological brains. However, the prevailing paradigm in AI views learning as only adjusting the strengths of synaptic connections in artificial brains, without any developmental mechanisms. Here it is contended that giving machines the ability to learn how to grow and adapt their artificial brains in response to the environment will enhance their learning abilities and allow them to continually improve performance during execution. I will develop and study a new class of such algorithms, in which an evolved neural developmental program is able to generate machines that learn, instead of relying on human-designed learning mechanisms. This project may also shed light on some of the most fundamental questions in neuroscience, such as the neural mechanisms underlying learning, their evolutionary origins, or the question of nature vs. nurture. The techniques developed here will allow adaptive machines to be grown instead of programmed and open new research avenues for AI.
Akronym | GROW |
---|---|
Status | Afsluttet |
Effektiv start/slut dato | 01/10/2019 → 28/02/2021 |
Finansiering
- Danmarks Frie Forskningsfond: 847.022,00 kr.