Designing the Future of Artifical Intelligence
True General Intelligence Current state-of-the-art neural networks employed in image recognition and decision making software require enormous, highly specialized datasets for their training. The result is a rigid AI that is only capable of fulfilling a single, well defined task. This restriction does not reflect the evergrowing challenges of the real world. As intelligent beings, we overcome difficulties by recollecting knowledge from many independent domains. When our understanding of the world reaches a limit, we explore it and bend its rules in order to independendly increase our knowledge. All without a pre-labeled database.
What we seek is a general artificial intelligence which is capable of the same autonomous solution-finding that humans are capable of since their infancy.
Current state-of-the-art neural networks employed in image recognition and decision making software require enormous, highly specialized datasets for their training. The result is a rigid AI that is only capable of fulfilling a single, well defined task. This restriction does not reflect the evergrowing challenges of the real world. As intelligent beings, we overcome difficulties by recollecting knowledge from many independent domains. When our understanding of the world reaches a limit, we explore it and bend its rules in order to independendly increase our knowledge. All without a pre-labeled database.
What we seek is a general artificial intelligence which is capable of the same autonomous solution-finding that humans are capable of since their infancy.
Realistic Neural Networks Instead of treating the mind as a black box with clearly defined inputs that are translated into outputs, we interpret the brain as a complex intertwined machine with an everchanging internal state. Our neural networks are modelled after real neurons and their connections, down to the electrochemical level. The neurons themselves hold an electrical charge that changes over time, allowing the network to adapt to new situations. This eliminates the traditional division between separate training and working phases, allowing the network to continuously learn while facing its challenges.
Instead of treating the mind as a black box with clearly defined inputs that are translated into outputs, we interpret the brain as a complex intertwined machine with an everchanging internal state. Our neural networks are modelled after real neurons and their connections, down to the electrochemical level. The neurons themselves hold an electrical charge that changes over time, allowing the network to adapt to new situations. This eliminates the traditional division between separate training and working phases, allowing the network to continuously learn while facing its challenges.
Genetic Evolution A neural network needs a way to build upon its success in order to constantly improve. We think that the best role model for a training method is the real world and the circumstances which lead to our species' own intellect. The blueprint of our simulated brains is encoded in an artificial DNA, which changes and improves the neural networks by Darwin's universal laws of evolution. The simulated genetic structure is modeled after the most recent discoveries in the young field of evolutionary developmental biology, preserving the inherent flexibility that gave birth to the countless forms of life we see every day.
A neural network needs a way to build upon its success in order to constantly improve. We think that the best role model for a training method is the real world and the circumstances which lead to our species' own intellect. The blueprint of our simulated brains is encoded in an artificial DNA, which changes and improves the neural networks by Darwin's universal laws of evolution. The simulated genetic structure is modeled after the most recent discoveries in the young field of evolutionary developmental biology, preserving the inherent flexibility that gave birth to the countless forms of life we see every day.
Simulation Based In our opinion, the primary factors behind humanities intellectual dominance are the abilities to adapt to environmental changes and to band knowledge together when confronted with problems that outscale one's own personal abilities. These qualities cannot be expressed in abstract mathematical terms. They can only be learned by being exposed to the world. This is why the training of our neural networks happens inside a simulated physical world. Because we programmed this simulation on top of our own in-house developed engine, we have the freedom to introduce arbitrary dynamic tasks that the simulated organisms need to overcome adaptively in order to propagate. This way, we can guide the ensuing evolution in whatever direction we need.
In our opinion, the primary factors behind humanities intellectual dominance are the abilities to adapt to environmental changes and to band knowledge together when confronted with problems that outscale one's own personal abilities. These qualities cannot be expressed in abstract mathematical terms. They can only be learned by being exposed to the world. This is why the training of our neural networks happens inside a simulated physical world. Because we programmed this simulation on top of our own in-house developed engine, we have the freedom to introduce arbitrary dynamic tasks that the simulated organisms need to overcome adaptively in order to propagate. This way, we can guide the ensuing evolution in whatever direction we need.