A team led by Professor Se-Bum Paik from the Department of Brain and Cognitive Sciences has employed AlexNET, a brain-inspired AI model, to understand the emergence of number sense in young animals even without training. Many animals have an intuitive sense for the arithmetic relationship between two quantities seen. While the mechanism for this perception of visual quantity is already well-understood, this is less so for young animals and even infants who have not had the necessary time to train this capacity. The study, published in August in Cell Reports under the title “Comparison of Visual Quantities in Untrained Neural Networks”, provides a new model for potentially understanding the mechanism for this ability.

The breakthrough in computer vision systems involving the use of convolutional neural networks about a decade ago resulted in the development of complex models like AlexNET. These systems were based on research into the visual cortex of cats and how their neurons collected and combined information. Convolutional neural networks apply the same principle to images, extracting information from a group of pixels at a time. It was this basic model, combined with increased processing power, that eventually produced the powerful recognition models we have today.

As a model for the untrained visual system of a young animal, the AlexNET model was initialized with random parameters for its neurons. Each “neuron” is a simple mathematical model of the function of an actual neuron, producing a scaled combination of its inputs as its output. The team supplied the model with stimuli in the form of black and white dots in its input. Upon monitoring the network's response to various inputs of white and black stimuli, they were able to identify particular regions in its output that responded strongly to changes in the proportion of white and black dots. Other regions that only responded to changes in the difference between the number of each type of dot were also found. 

Notably, difference-sensitive neurons have not been found in animal experiments, while the model suggests that they are likely to exist. According to the paper's first author Hyeonsu Lee, a PhD candidate from the Department of Bio and Brain Engineering, computational models like this can work alongside experiments to improve the efficiency of research in neuroscience by providing a direction to work toward. He believes the results could potentially form the basis for understanding the initiation of cognitive functions in the brain. 

Lee suggested that the emergence of useful behavior in untrained networks could potentially be applied, by developing "training algorithms that harness these innate tunings", which could result in cheaper training. Amidst a global chip shortage, algorithmic improvements like these could make accessible models that had previously been economically impractical to train.