All topographic neural networks mimic the human visual system more closely

Deep learning models such as convolutional neural networks (CNNS) and recurrent neural networks (RNNs) are designed to partially emulate the functionality and structure of biological neural networks. As a result, in addition to tackling a variety of real-world computational problems, it also helps neuroscientists and psychologists to better understand the foundations of specific sensory or cognitive processes.

Researchers at Osnabrück University, Freie Universität Berlin and other laboratories have recently developed a new class of artificial neural networks (ANNs) that can better mimic human visual systems than CNNs and other existing deep learning algorithms. Their newly proposed visual system-inspired computational techniques are called All Topographic Neural Networks (All-TNNS) and featured in a paper published in Natural human behavior.

“Previously, it was the most powerful model for the brain to understand how visual information was derived from AI vision models,” Dr. Tim Kietzmann, senior author of The Paper, told Tech Xplore.

“These are often convolutional in nature. Machine learning hacks where the corresponding neural network can search for the same functionality anywhere in the visual input. This approach is very powerful. Learning in one place in space is something the brain cannot do (the brain cannot copy.”

In addition to performing several actions that the primate brain cannot perform, CNNS organizes information that is different from biological neural networks. In contrast to CNNS, the brain is organized in the retina. In other words, visual signals travel from the retina to the visual cortex (a region of the outer layer of the brain where visual information is known to be processed).

“The brain also shows a systematic relationship between the types of traits that are responding to and where they are looking for them,” Kietzmann said.

“This interaction of space and features along the cortical surface is an important aspect of visual processing, but as mentioned above, this feature is not considered in machine learning. To solve this drawback, we have the function selectivity closer to “cortical sheets.”

Most computational approaches commonly used to model the way human visual systems process natural images rely on deep neural networks (DNNSs), such as CNNSs. These are powerful models that can be trained to classify visual data such as brain imaging scans and to identify specific objects in images.

“The problem with these models is that they are often quite far from biology and, despite being more powerful, have stopped the new ML models being better models of visual processing in the brain (relationships that were true in the past),” explained Kietzmann.

“In a series of papers, my lab shows how to modify the ML model to make the ML model a better model of biology. For example, by training better image datasets, including recurrent connectivity in network architectures, by considering the task of training the model, and more recently, by considering the brain alignment along the cortical surface.”

Kietzmann and his colleagues demonstrated that the new model developed based on (All-TNNS) reflects more closely the human visual system than CNNs and other DNNs. This is because it replicates the principles that underpin the tissues of the visual cortex, but also better captures human behavior patterns than previously developed models.

In the future, all TNNs may be used to conduct neuroscience and psychology research, shedding new light on the neural foundations of the human visual system. For example, it helps us to better understand how feature selectivity arrangements across the cortex, also known as topography, affect human perception and behavior.

“Currently, the terrain network is rich in parameters, and we are looking to improve our training to be more efficient in terms of task performance compared to our convolutional counterparts,” Kietzmann added.

“In addition, models need to be manipulated at present towards smooth feature selectivity across space, a key feature of cortical topography. However, biology may have developed an implicit mechanism that smooths cortical selectivity.

©2025 Science X Network