The aim of artificial intelligence research is to develop a machine capable of undertaking any cognitive task that the human brain can perform, writes Simon Stringer at the Oxford Laboratory for Theoretical Neuroscience and Artificial Intelligence. The intellectual flexibility of this artificial general intelligence (AGI) would surpass even the best AIs available today, whose performance is limited to specific spheres such as playing games or recognising images.

We need to pay close attention to the architecture of the brain, and especially how real neurons communicate, if we are going to understand the biological foundations of consciousness and develop AGI machines.

What is artificial general intelligence (AGI)?

© Science Photo Library

AGI would replicate brain function in computers. It would create a machine that can perceive and interpret its environment at every spatial scale, use this rich sensory information to learn causal relationships in the world, and then process its knowledge to produce intelligent behaviour in any complex real-world environment.

Why does AGI matter?

AGI will open up a much broader range of applications in machine intelligence and robotics across industrial sectors such as social care, manufacturing and defence. For example, smart vision could be used to spot when an elderly person has fallen at home. This is still a big challenge for current AIs, while human vision can solve the task effortlessly. Brain-inspired neural networks are also being applied to military problems — they could, for example, be mounted on autonomous drones to spot enemy tanks and other military vehicles.

Limitations of today’s AI

Google’s deep learning neural networks learned to recognise cats after training on millions of random images from the internet. However, these artificial networks do not “see” the cat in the same rich semantic way as the brain. Instead, they simply respond on a “yes/no” basis depending on whether a cat is present or not. Computer experts can easily trick such networks into making mistakes, by manipulating images in a way that AGI would spot immediately.

Lessons from the brain

In contrast, when humans look at an image of a cat, we see the hierarchy of features that comprise the animal such as the mouth and head, including how these features are related to each other. Such semantic…

Continue reading: