New leap in AI—All-TNNs mimic how the brain sees.

A new class of artificial neural networks, known as All-Topographic Neural Networks (All-TNNs), has been developed to more closely mirror the functioning of the human visual system. Unlike traditional models such as convolutional neural networks (CNNs), which apply the same feature detectors uniformly across an entire visual field, All-TNNs adopt a more biologically realistic approach. They incorporate a retinotopic and topographic organization, meaning features are not only detected based on their presence but also on where they appear, emulating the spatial processing seen in the brain’s visual cortex.

This spatially structured design contrasts sharply with the "copy-paste" strategy used in CNNs, where learned features in one part of the image are generalized across all locations. In the human brain, such uniform transfer does not occur. Instead, visual processing depends heavily on both feature type and location, with neurons tuned to specific regions and characteristics. All-TNNs simulate this by aligning feature detectors across a two-dimensional cortical-like surface, ensuring that neighboring regions are responsive to similar features, while more distant regions exhibit variation.

This architecture allows All-TNNs to better capture human behavioral patterns in visual tasks, offering improved alignment with how visual information is naturally processed. While deep neural networks have long been used to model aspects of visual perception, recent advancements in AI have often led to architectures that diverge from biological plausibility. All-TNNs help close this gap by reintroducing key organizational principles found in the brain.

In addition to offering greater biological realism, All-TNNs present opportunities for new research in neuroscience and psychology. They could help uncover how the spatial arrangement of feature sensitivity—also called cortical topography—influences perception and cognition. Ongoing efforts are focused on improving training efficiency and achieving smoother transitions of feature selectivity across the artificial cortex, as observed in the brain. Exploring how biological systems naturally achieve this smoothness remains a compelling research direction.