Artificial intelligence models used to explain vision can predict patterns in recorded brain activity, yet those patterns do not predict many of the models’ internal features. The one-way match raises a basic question for neuroscience: when a model mimics a brain response, does it reproduce the process that produced it?
A ScienceDaily report, based on materials provided by York University and written by Sandra McLean, describes a study by Sabine Muzellec and Kohitij Kar. The report identifies Kar as the study’s senior author. The paper, “Reverse predictivity for bidirectional comparison of neural networks and biological brains,” was published in Nature Machine Intelligence.
Researchers have used artificial neural networks, or ANNs, to study vision because these computer systems can perform tasks such as recognizing objects and reproduce patterns of brain activity. The York team tested whether that resemblance held up when they reversed the usual comparison. Their method, called reverse predictivity testing, asked whether recorded neuronal activity could predict features inside the AI models, as well as the models could predict the recorded neural responses.
The tests used 1,320 natural photographs and realistic synthetic images of 10 familiar subjects, including animals, faces and objects, shown against varied scenes. The team also tested 300 images of the same subjects in styles such as outlines, drawings, simplified representations and artistic interpretations. The report does not detail the number or architecture of the AI models tested.
The models generally predicted the responses of recorded neurons. In the reverse direction, however, those neural responses were much less effective at predicting many of the models’ internal features. The report says the researchers did not observe the same problem when comparing neuronal activity across different brains. That comparison suggests the asymmetry they found is not simply an inevitable feature of comparing neural responses with one another.
The result does not mean that AI models are useless for neuroscience, or that every part of a model differs from the brain. It shows that a model’s ability to predict a measured brain response is not, by itself, evidence that its internal processes match biological ones. The report says the parts of models that align with the brain also do better at predicting real human behavior, but it gives no specific measures for that relationship.
That distinction matters because researchers use AI models to study behavior and the nervous system. If a model produces a brain-like response through different internal strategies, researchers could mistake a successful prediction for a faithful explanation of how perception works. The report says similar models are also used to study hearing, language and movement, but the study described here tested vision; it does not establish whether the same asymmetry occurs in those fields.
The team presents reverse predictivity as a diagnostic that can help researchers identify which parts of an ANN correspond to real neural activity. York University’s report says the researchers released a public testing toolkit for evaluating models. The work therefore offers a way to scrutinize claims of brain likeness, while leaving open how broadly the results apply across models, datasets and other areas of neuroscience.