According to a new study from the University of Oxford, the human brain has a fundamentally different and more efficient way of learning compared to machines.
Fundamental Learning Process
One of learning's fundamental processes is something known as credit assignment. When one commits a mistake, credit assignment attempts to figure out where the error was introduced in the information-processing pipeline.
The majority of modern AI has artificial neural networks. These are different layers of nodes or neurons that are similar to what can be found in the human brain. When artificial intelligence performs a mistake, it modifies the links between the neurons, which is also called adjusting the weights. It does so in order to fine-tune the process of decision-making until it obtains a right answer.
The process is known as backpropagation, since the errors end up propagating backwards through the neural networks of the AI.
Until just recently, several researchers believed that this was also how biological networks of brain neurons work in order to learn from novel experiences.
Human Brain vs. AI
In the new study "Inferring neural activity before plasticity as a foundation for learning beyond backpropagation," the researchers note that backpropagation has powered significant AI advancements since its inception. It has also earned quite a dominant place in understanding how the brain learns. They also note, however, that the brain is superior compared to AIs that use backpropagation in various ways.
Though AI has been seen to perform better than humans in various tasks such as hiring staff or creative thinking, it takes quite a while for AI to learn how to execute such things.
Humans have the capacity to learn from just one instance of a new experience. However, AI requires up to hundreds or even thousands of exposure times. Moreover, when humans learn something novel, it does not interfere with what humans already know. The opposite is the case for AI.
With such proof at hand, the researchers examined various equation sets that describe the changes in brain neurons. Upon simulating information processing methods, they discovered that this was an entirely different way of learning compared to backpropagation. They call this prospective configuration.
In contrast to AI wherein interneural connections are modified, neuron activity alters in order to better predict results. Weights are then adjusted in order for the new pattern to be matched.
Though this may not appear to be a stark difference, the impacts are actually quite significant. The researchers explained such differences and effects using a hunting bear illustration.
Moreover, the researchers demonstrated, using computer simulations, that models using prospective configuration are capable of learning faster and more effectively compared to AI neural networks. This was the case for tasks that humans and animals in nature may typically face.
While prospective configuration appears to be a more efficient learning method, scientists say that computers are not capable of using such a kind of system. Dr. Yuhang Song, the study's first author, explains that the simulation of this approach on current computers is slow due to how they fundamentally operate differently compared to the biological brain.
However, there is a possibility of developing new computers that are able to use the approach. Dr. Song notes that a new computer type or brain-inspired hardware must be developed for prospective configuration to be implemented with minimal energy use.
Professor Rafal Bogacz, the lead researcher, also notes that a big knowledge gap is currently present between reality and theory. Professor Bogacz explains that at present, there is a big gap between models that do prospective configuration and current knowledge regarding brain network anatomy. The professor adds that further research may aim to bridge this gap.
RELATED ARTICLE: Self-Supervised Learning Artificial Intelligence Resembles How the Human Brain Functions
Check out more news and information on Tech & Innovation in Science Times.