Understanding the neural code requires that we understand how neurons interpret our environment. The visual cortex relies on a hieararchy of processing steps to infer how a visual scene is composed. In this hierarchy neurons are sensitive to more and more complex constituents of the visual environment while in parallel develop invariance towards other features. AI helps to precisely characterize the steps in this multi-step process as it provides support what constituents to look for when inspecting the neural code.
The dominant AI methods to reverse-engineer the neural code are deep discriminative models that are trained to interpret images by inferring object categories and this is achieved with a learning paradigm called supervision, where each input image is paired with the desired output of the network. However, the biological brain is capable to face a much more demanding setting: being able to infer more properties of an image (for example the pose of an animal or its current level of hunger) and doing this without supervision.
We highlight that there is an extra asset in the sleeve of a biological neural network compared to deep discriminative models: while these image models operate solely with a sequential processing scheme, so called feed-forward processing, the biological brain relies on back-and-forth information flows. This is implemented by top-down connections between processing stages. This insight motivates a departure from discriminative models. Turning to deep generative models, in our paper we show that these heavily rely on backwards flow of information. When adapting this image model to the stimuli occurring in our natural environment, the top-down connections developing in the deep generative model predict the properties of top-down influences in the brain of primates in unprecedented detail and without the need to adjust model parameters specifically to fit the brain signals.
Through developing brain-compatible deep generative models, AI provides specific insights into how the details neural code can be understood in ever higher levels of detail.
Image caption:
A processing pipeline implies a one-way street for information flow. Using brain-inspired computing, deep generative models reveal a backward flow of information. In the visual cortex of non-human primates. Image credit: AI-generated image (OpenAI, ChatGPT, GPT-5)
Related publication:
Ferenc Csikor, Balázs Meszéna, Katalin Ócsai & Gergő Orbán (2025) Top-down perceptual inference shaping the activity of early visual cortex. Nature Communications, doi:10.1038/s41467-025-64967-x