Artificial intelligence systems built with biologically inspired designs may begin to resemble human brain activity even before they are trained on data, according to new research from Johns Hopkins University. This research suggests that how AI is structured may be just as important as the amount of data it processes.
The survey results are nature machine intelligencewe will challenge the dominant strategy in AI development. This research highlights the value of starting with a brain-like architectural foundation, rather than relying on months of training, massive datasets, and vast amounts of computing power.
Rethink your data-heavy approach to AI
“The way the AI field is progressing right now is to feed models with massive amounts of data and build out computing resources the size of a small city. You have to spend hundreds of billions of dollars to do that. Humans, on the other hand, learn to see using very little data,” said lead author Mick Bonner, assistant professor of cognitive science at Johns Hopkins University. “There may be good reasons why evolution converged on this design. Our study suggests that a more brain-like architectural design puts AI systems at a very advantageous starting point.”
Bonner and his colleagues aimed to test whether architecture alone could give an AI system a more human-like starting point without relying on extensive training.
Comparison of common AI architectures
The research team focused on three main types of neural network designs commonly used in modern AI systems: transformers, fully connected networks, and convolutional neural networks.
They iteratively tweaked these designs to create dozens of different artificial neural networks. None of the models are pre-trained. The researchers then showed the untrained system images of objects, people, and animals and compared their internal activity to the brain responses of humans and nonhuman primates viewing the same images.
Why convolutional networks stand out
Increasing the number of transformers and artificial neurons in a fully connected network resulted in little meaningful change. However, similar adjustments to the convolutional neural network resulted in activity patterns that more closely match those seen in the human brain.
According to the researchers, these untrained convolutional models performed on par with traditional AI systems, which typically require exposure to millions or even billions of images. The results suggest that architecture plays a larger role in shaping brain-like behavior than previously thought.
A faster path to smarter AI
“If training on large amounts of data is really the key element, then there is no way to achieve a brain-like AI system just by changing the architecture,” Bonner said. “This means that by starting with the right blueprint and perhaps incorporating other insights from biology, we have the potential to dramatically accelerate learning in AI systems.”
The team is currently exploring simple learning methods inspired by biology. This could lead to a new generation of deep learning frameworks, potentially making AI systems faster, more efficient, and less dependent on large datasets.
