Summary: New research using artificial neural networks shows that number sense can emerge spontaneously from the visual system, without any prior counting experience.
Source: University of Tübingen
Humans and many animals possess an innate “number sense” — the ability to estimate how many objects are present in a scene. Neuroscientists attribute this ability to so-called number neurons, cells that respond selectively to particular numerosities and that have been observed in both human and animal brains. A longstanding question has been whether these number neurons arise directly from visual processing and, if so, how that emergence occurs. A research team led by Professor Andreas Nieder at the Institute of Neurobiology, University of Tübingen, explored this question by training and probing a biologically inspired artificial neural network. Their findings, published in Science Advances, indicate that numerosity representations can develop spontaneously within a visual system trained only for object recognition, without any explicit training to count.
The study began by training a deep convolutional neural network to recognize a wide variety of photographed objects — for example, tennis balls, necklaces, spiders, and dogs. The network architecture was modeled on early stages of the human visual cortex, where neurons are organized hierarchically to extract progressively more abstract visual features. The network was trained on 1.2 million labeled images spanning one thousand object categories. After training, the model achieved high accuracy when classifying thousands of new images, demonstrating robust object-recognition capabilities.
Number sense emerges from visual feature processing
The artificial network consists of two functional parts: the first extracts and transforms visual features into abstract internal representations; the second maps those representations to categorical outputs with probabilistic classification. To test whether numerosity-sensitive units could arise from visual feature processing alone, the researchers separated the feature-extraction portion from the classification head. They then presented the feature extractor not with photographs but with simple dot arrays containing between one and thirty dots. The dot displays varied systematically in size, density, and spatial arrangement across trials so that any neuron responding to number would have to do so independently of those other visual attributes.
Analysis of the network’s internal units revealed that nearly ten percent of the artificial neurons became selectively tuned to specific numerosities — for example, preferentially responding to displays of three dots, five dots, or other particular counts — even though the network had never been trained to discriminate number. These numerosity-tuned units emerged spontaneously as a property of the visual feature representations learned during object-recognition training.
The properties of these artificial numerosity units closely resembled those of number neurons recorded in animals and humans. The model’s number discrimination performance followed the same behavioral signature found in biological organisms, including variability consistent with the Weber–Fechner law: discrimination accuracy declined as the numerical difference between two sets became proportionally smaller. This parallel suggests that mechanisms inherent to visual object recognition are sufficient to explain the spontaneous emergence of number-sensitive units.
Professor Nieder emphasizes that number sense need not depend on a highly specialized brain area dedicated solely to counting. Instead, it can be regarded as an emergent property of neural networks shaped by visual experience and the demands of object recognition. This perspective helps explain why newborn infants and untrained wild animals demonstrate basic numerosity estimation: the visual system’s architecture and its role in extracting object features naturally give rise to neurons that encode number.
Source:
University of Tübingen
Media contact:
Andreas Nieder – University of Tübingen
Image source:
The image is in the public domain.
Original research (open access):
“Number detectors spontaneously emerge in a deep neural network designed for visual object recognition.” Khaled Nasr, Pooja Viswanathan, Andreas Nieder. Science Advances. doi: 10.1126/sciadv.aav7903
Abstract
Number detectors spontaneously emerge in a deep neural network designed for visual object recognition
Humans and many animals possess an innate ability to estimate numerosity — the number of visual items in a set. This work shows that units tuned to abstract numerosity, analogous to biological number neurons, can emerge spontaneously in a biologically inspired deep neural network that was trained only for visual object recognition. These numerosity-tuned units support the network’s number discrimination behavior, which mirrors key characteristics of human and animal number perception as predicted by the Weber–Fechner law. The findings demonstrate that the number sense can arise from mechanisms intrinsic to the visual system, without explicit training to count.