Researchers have unveiled the All-Topographic Neural Network (All-TNN), a novel neural network demonstrating more human-like image perception compared to prevalent Convolutional Neural Networks (CNNs).

The key distinction of All-TNN lies in its architecture. Unlike CNNs, which replicate feature detectors, All-TNN possesses a unique topographic structure where each spatial location has its own set of trainable parameters. This mimics the human brain's approach, which does not copy knowledge. During training, a "smoothness constraint" is employed, enabling neighboring neurons to learn similar, but not identical, features.

In object recognition tests, All-TNN exhibited a threefold stronger correlation with human perception than CNNs. While its classification accuracy currently trails that of CNNs, All-TNN demonstrates significant superiority in energy efficiency: despite being 13 times larger, it consumes over 10 times less energy. This is achieved by concentrating resources on the most informative parts of an image.

The authors emphasize that their objective was to create an architecture that advances the understanding of the working principles of both artificial and human intelligence, rather than solely focusing on energy efficiency. The development of All-TNN represents a crucial step toward creating neural networks that simulate human-like behavior.

ORIENT