A recent study has concluded that internal noise in neural networks does not necessarily impose a fixed ceiling on the amount of information the brain can represent, contrary to the assumptions of mathematical models and studies over about 30 years.
Study reassesses the ceiling on neural information
The findings, published in Science Advances, showed that the capacity of neural networks in mammalian brains can continue to rise as the number of participating cells increases, although each additional cell makes a progressively smaller contribution. The study addresses a fundamental question in neuroscience: how the activity of thousands of neurons turns what the eye detects into a clear and intelligible image, such as perceiving a branch moving in the wind or distinguishing fine details in a painting.
The question is whether neural networks reach a point beyond which adding new cells is no longer useful, or whether a larger network can always represent more information. Hideaki Shimazaki, an associate professor at Kyoto University's Graduate School of Informatics, said resolving the issue required real-world experimental data.
He said technologies available in recent years have become capable of recording the activity of 1,000 or more neurons simultaneously, allowing researchers to track directly how information increases as the number of cells involved in processing rises.
The pessimistic assumption that prevailed in many earlier studies was based on the idea that neurons are not independent, isolated computing units, but interconnected biological units operating within a chemical environment and affecting one another.
Under this view, random fluctuations, or noise, affecting one cell can spread to neighbouring cells and accumulate to the point where adding more cells can no longer increase the available information. To test this assumption, the research team focused on the primary visual cortex in mammals and reanalysed detailed experimental records of the activity of about 20,000 neurons in the brains of five mice.
The visual cortex in mice is used as a biological model for studying the general mathematical principles governing image processing in mammalian brains. Shimazaki explained that a neuron in the visual cortex does not produce the same response every time the animal is shown the same image. The variation between responses is called “neural noise”.
Shared noise slows information but does not block it
By contrast, a slight change in an image alters the average activity of neurons, and this change forms the signal the brain uses to distinguish between two images. Mathematical analyses showed that fluctuations shared between cells do not accumulate into a random mass that blocks the visual signal, as previously believed.
Shimazaki said data from all five mice placed them in the category of “continued information growth”, explaining that shared noise slows the growth of information but does not stop it as long as the rules observed by the researchers continue to operate. According to the study, if a slight change in the orientation of visual lines represents a signal moving in a specific direction, noise spreads through the space of cell activity according to a system known mathematically as “scale invariance”.
The concept means that something retains the same pattern regardless of differences in its size. Shimazaki said the noise pattern recorded in small and large groups of neurons takes the same form after the calculations are adjusted to reflect the size of each group.
This pattern helped the researchers explain why information continues to increase despite correlated fluctuations among large numbers of cells. According to the analysis, brain noise can be divided into a limited number of strong, high-intensity patterns and a very long list of weak, low-intensity patterns. Earlier studies focused mainly on whether the strong patterns could block the signal the brain needs.
But after identifying the precise condition under which information could reach its maximum, the team found that weak patterns must not be overlooked. Shimazaki explained that the signal is barely blocked when a weak pattern is present, meaning that a small part of the signal in this range can carry a large amount of information.
Weak patterns reveal signal pathways
He added that the large number of quiet patterns makes their combined effect capable of determining the outcome, noting that the study demonstrated for the first time the need to examine the full range of noise rather than focusing only on its strongest components. The flow of neural data can be compared with a multi-lane motorway network.
The main routes may experience congestion and increasing fluctuations, but the direction of the large noise in them does not exactly match the direction of the visual signal the brain is trying to detect and distinguish. Alongside those routes, the network contains thousands of quiet side lanes where noise falls to limited levels and which are geometrically perpendicular to the direction of the signal.
This geometric separation allows the quiet pathways to remain open for the signal, enabling the neural system to continue increasing the accuracy of sensory discrimination cumulatively as new neurons join the network. The findings do not mean that the brain has an unlimited capacity to process information. The researchers stress that processing capacity is in fact subject to physical and biological constraints.
Shimazaki said what the visual system can use is ultimately determined by the amount of information coming from the eye, because the retina contains a limited number of cells and transmits only a limited amount of data. The visual cortex itself also contains a limited number of neurons.
More specifically, the study indicates that internal noise measured in the primary visual cortex of mice does not by itself impose an information ceiling, provided the observed rules continue to apply. Each additional cell contributes slightly less than the preceding one, in a pattern resembling a staircase whose steps become shorter but which continues to rise, rather than reaching an absolute barrier that prevents any further increase.
Study findings open applications for artificial intelligence
The findings could have applications beyond theoretical neuroscience, offering a concept that artificial intelligence engineers and developers of artificial neural networks could use when designing machine-learning algorithms that resist internal noise while remaining efficiently scalable.
They could also contribute to the development of brain–computer interfaces and neural prostheses by training them to extract sensory patterns from quiet pathways and avoid getting caught in electrical noise.
