Artificial intelligence (AI) can imitate brain activity even before learning from data if built like the human brain, according to researchers at Johns Hopkins University. They question the usual way of creating AI, which relies on feeding huge amounts of data into models through deep learning - a process where AI improves by analyzing patterns in data. This traditional method takes months, costs billions of dollars, and uses massive energy.
The researchers note that humans learn to see with little data, unlike current AI that needs vast resources. Evolution might have shaped the brain's design for efficiency, and copying it could give AI a strong start.
The study examined three common AI blueprints: transformers, fully connected networks, and convolutional networks, which process images by focusing on local patterns.
Architectural changes drive brain-like results
The researchers altered these designs to create many unique artificial neural networks. Without training, they showed these networks images of objects, people, and animals, then compared their reactions to brain scans from humans and primates viewing the same images.
Adding more artificial neurons did little for transformers or fully connected networks. But for convolutional networks, this tweak created activity patterns closer to the brain's. These untrained networks performed as well as AI trained on millions of images, showing that the blueprint, or architecture, matters more than once thought.
The researchers argue that if data training were key, tweaks alone could not achieve brain-like results. Starting with the right design, plus biology insights, might speed up AI learning greatly.
Looking ahead, the researchers are creating simple learning rules based on biology for a new AI framework. This could make AI more efficient, reducing costs and energy use while advancing fields like robotics and medicine.
The methods and results of this study are described in a paper published in Nature Machine Intelligence.