An international collaboration between researchers from the RIKEN Center for Brain Science (CBS) in Japan, the University of Tokyo, and University College London has made a significant breakthrough in understanding the self-organization of neurons during the learning process. The study reveals that the behavior of neural networks as they learn can be accurately predicted by a mathematical theory known as the free energy principle.
This discovery has implications for the development of animal-like artificial intelligence and for comprehending cases of impaired learning. The study was published on August 7th in Nature Communications. In the intricate landscape of the human brain, neurons continuously reorganize themselves as they learn to distinguish between various sources of incoming information such as voices, faces, or smells.
This adaptive process involves the alteration of the connections’ strength between neurons, serving as the fundamental basis for all forms of learning within the brain. A team led by Takuya Isomura from RIKEN CBS and international colleagues postulated that the self-organization of neural networks during this learning process adheres to the mathematical principles of the free energy principle. This hypothesis was put to the test using neurons extracted from rat embryos and cultivated in a dish with a grid of miniature electrodes beneath them.
The experimental procedure simulated the learning process by stimulating the neural network through the grid of electrodes, emulating the phenomenon where neurons respond distinctly to different sources of sensory input after learning. This was replicated by stimulating neurons in a specific pattern that combined two separate hidden sources of input.
After undergoing 100 rounds of training, the neurons exhibited automatic selectivity – some responded vigorously to source #1 and weakly to source #2, while others responded inversely. The disruption of the learning process occurred when drugs that either heightened or lowered neuron excitability were introduced to the culture prior to the experiment. This confirmed that the cultured neurons behaved similarly to the neurons within a functioning brain.
The free energy principle stipulates that such self-organization follows a pattern that consistently minimizes the free energy within the system. To ascertain whether this principle governed the learning process of neural networks, the researchers constructed a predictive model based on the actual neural data. This model was then tested by feeding it data from the initial 10 electrode training sessions, which was then utilized to predict the subsequent 90 sessions. Remarkably, the model accurately forecasted the responses of neurons and the strength of connectivity between them at each stage.
This implies that having knowledge of the neurons’ initial state is sufficient to determine how the network will evolve as learning transpires over time. Takuya Isomura highlights the significance of their findings, stating, “Our results suggest that the free-energy principle is the self-organizing principle of biological neural networks. It predicted how learning occurred upon receiving particular sensory inputs and how it was disrupted by alterations in network excitability induced by drugs.”
Furthermore, Isomura envisions broader applications for their research, indicating that their technique could potentially be used to model the mechanisms underlying psychiatric disorders and the effects of various drugs, such as anxiety-reducing medications and psychedelics. The research’s implications extend even further, offering insights into creating next-generation artificial intelligence models that can learn in a manner resembling real neural networks.
The collaboration of international researchers has unveiled a groundbreaking understanding of neural network self-organization during the learning process. Their work demonstrates that the free energy principle accurately predicts this process, shedding light on potential applications in fields like artificial intelligence development and the study of neural disorders. This discovery deepens our comprehension of the intricate workings of the human brain and its potential applications across various scientific domains.
