- New research from Johns Hopkins University shows that AI can exhibit human-brain-like behavior even without being trained on any data.
- This discovery challenges the current dominant approach that relies on massive datasets, months of training, and hundreds of billions of dollars in computing infrastructure.
- Scientists are focusing on modifying AI architecture to more closely resemble the biological brain, rather than “stuffing” it with more data.
- The study, published in Nature Machine Intelligence, emphasizes that system structure is as vital as the amount of input data.
- The research team compared three popular AI architectures: transformers, fully connected networks, and convolutional neural networks (CNNs).
- Dozens of different models were created without pre-training, then shown images of objects, people, and animals.
- The internal activities of these models were compared to the brain responses of humans and primates observing the same images.
- Increasing the number of neurons in transformers and fully connected networks made almost no significant difference.
- Conversely, untrained CNNs produced activity patterns remarkably close to the human brain.
- The effectiveness of these models is comparable to traditional AI that requires training on millions or billions of images.
- Results suggest that AI architecture can dictate “brain-like” behavior more powerfully than data itself.
- The team is exploring simple, bio-inspired learning methods to create more efficient deep learning frameworks.
📌 Conclusion: New research from Johns Hopkins University demonstrates that AI can display brain-like behavior even without prior training data. Scientists are shifting focus toward bio-inspired AI architectures instead of data-heavy scaling. This not only opens new directions for AI but also raises significant questions about current generative AI models that depend on massive data.
