IFISC researchers discover how collective learning emerges in neural networks
Nov. 13, 2024
- A study published in Physical Review Research reveals that, by coupling different neural networks, they can learn from each other without the need to exchange data, and collective learning emerges abruptly, through phase transitions.
- The work developed by IFISC scientists - with implications for the security of artificial intelligence systems - uses methods from physics to theoretically and practically reveal the emergence of collective phenomena in swarms of neural networks, thus deepening our theoretical understanding of artificial intelligence.
A team of researchers from the Institute of Cross-disciplinary Physics and Complex Systems (IFISC, UIB-CSIC) has developed an effective theory that predicts and explains how collective learning emerges in systems of deep neural networks placed in interaction. The study, published in the journal Physical Review Research, shows that this phenomenon arises through phase transitions, whose properties in turn depend on the depth of the neural networks.
The team, formed by Lluís Arola-Fernández and Lucas Lacasa, designed a system of several independent neural networks, where each neural network is a toy brain that is trained to recognize a single class of data, such as a particular digit in a set of images.
The researchers wondered what happens when several such brains are coupled together. “We wanted to understand if, by a sort of collective intelligence, by putting the brains to interact without sharing data, they could recognize images they had never seen before,” says Lucas Lacasa, IFISC researcher and co-author of the study. “We also wanted to understand whether this eventual transition to collective intelligence would happen gradually or abruptly and spontaneously,” explains Lluís Arola-Fernández, also at IFISC and co-author of the study.
The researchers developed a theory that predicts that this phenomenon occurs through what is known as a phase transition, determined by the intensity of the interaction between the brains. “Each network is trained to identify only one type of image, such as the number one or two, being very effective at detecting its own digit but ineffective with others. However, when interacting with each other, they manage to recognize all numbers,” explains Arola-Fernandez.
The study uses statistical physics techniques to describe how interactions between networks can lead to emergent behavior. “We have found that, surprisingly, the transition to the state of collective intelligence happens in a manner analogous to a type of spontaneous magnetization that occurs in certain materials, which points to a deep relationship between artificial intelligence and physics,” adds Arola-Fernandez. “Furthermore, our theory predicts that the depth of each of the brains (the number of layers in the neural network) is a parameter that radically changes the physics we observe.” This suggests that deep learning architectures could play a crucial role in how learning occurs in decentralized environments, with applications in areas as diverse as education or healthcare.
In their experiments, they validated this theory using datasets such as MNIST and CIFAR-10, two collections of images that are commonly used to train machine learning algorithms. The results are promising: “We have observed that individual networks, trained on a selected dataset, can fully generalize to unseen classes of data when this collective learning phase emerges,” says Lacasa. “This abrupt collective learning that we observed could also occur in social systems,” speculates Lluís, ”since the effective theory does not rely on the details of individual brains to explain the collective effect.”
The two researchers consider that the work also has disturbing ethical implications for the famous alignment problem in artificial intelligence: “This work is a simple proof-of-concept demonstrating that AI systems can suddenly acquire new capabilities for which they were neither trained nor designed, as soon as these systems enter into interaction. In our case, these new capabilities seem advantageous and harmless, but this will not necessarily always be the case,” they say.
Lluís Arola-Fernández and Lucas Lacasa, Effective theory of collective deep learning, Physical Review Research, 6, L042040, https://doi.org/10.1103/PhysRevResearch.6.L042040
