Artificial Intelligence Shows Human-like "Number Sense"

Source: GDJ/Pixabay

At the intersection of psychology and neuroscience is a phenomena called “number sense”—an intuition of the magnitude and relationships of numbers that is found in humans. In children and adolescents, research studies have shown that there is a relationship between number sense and mathematical aptitude. Less than a week ago, a team of researchers have achieved a significant milestone in artificial intelligence (AI) by extending number sense to machines.

On May 8, 2019, Andreas Nieder, Pooja Viswanathan, and Khaled Nasr published in Science Advances their landmark study demonstrating machine intelligence gaining number sense spontaneously on its own.

Not only is number sense is found in humans, but also in insects and animals. Honeybees use numerical cues to navigate. The ability to discriminate quantities in the animal kingdom is not unusual. Rhesus monkeys understand the concepts of addition, subtraction, “greater than” and “less than.” North Island robins of New Zealand rely on quantity discrimination for food retrieval and storage strategies to minimize spoilage losses. Now this decidedly biological instinct is no longer limited to only living organisms.

In the study, the research team trained a type of deep neural network, specifically a Heterogeneous Convolutional Neural Network (HCNN), to classify objects in natural images from the ImageNet dataset that contained around 1.2 million images in 1000 categories. In the HCNN, layers zero through 13 were for feature extraction, and layers 14 and 15 were for classification.

Once the network was trained, the team tested the object classification with 50,000 different images that were entirely new to the network, and not part of the training dataset. By extracting features from the training data, the artificial neural network was able to perform successful classifications such as distinguishing a wolf spider from other arthropods, a box turtle from a mud turtle, and a miniature schnauzer from a standard one.

Can this network, trained on object classification with natural images, spontaneously demonstrate number sense? To find out, the researchers discarded the classification network and used only the feature extraction network for the next part of the study. The team presented 336 images consisting of one to 30 white dots patterns on a black background to the feature extraction network, and observed the network’s neuronal response firings.

The team discovered that the network’s neurons acted in a similar manner as those in the brains of monkeys who previously viewed the same dot images. In other words, the artificial neural network neurons “fired” in a similar manner as neurons in the visual cortex of biological brains. The researchers wrote, “The responses of numerosity-selective units exhibited a clear tuning pattern,” which was “virtually identical to those of real neurons.” Number-sensitive artificial neurons, like the biological neurons in monkeys, would fire preferentially when viewing a certain number of dots.

“These findings suggest that the spontaneous emergence of the number sense is based on mechanisms inherent to the visual system,” wrote the researchers. “The workings of the visual system seem to be sufficient to arrive at a visual sense of number. Numerosity selectivity can emerge simply as a by-product of exposure to natural visual stimuli, without requiring any explicit training for numerosity estimation.”

The researchers further wrote, “Beyond providing an explanation of the neuroscience of the number sense, our approach also highlights how artificial neural networks give rise to unexpected feature selectivity that helps to understand emergent properties of the brain.”

Copyright © 2019 Cami Rosso All rights reserved.