Cosa è successo
York University scientists published a study in Nature Machine Intelligence demonstrating that artificial neural networks can predict brain‑cell responses to visual stimuli without processing images in the same way the human visual system does. They introduced a “reverse predictivity testing” method that evaluates whether brain activity can predict internal model features, finding that while AI models reliably forecast neuronal responses, the reverse is far less effective. The team also released a public toolkit so other researchers can assess how closely their models align with neural activity.
The York University team examined 1,320 natural photographs and 300 stylized versions (outlines, drawings, artistic renditions) shown to ten human subjects while recording neuronal activity. They compared the recorded responses to internal activations of several artificial neural networks trained for object recognition.
Using their reverse predictivity testing, they first confirmed that AI models could predict neuronal responses, replicating prior work. Then they attempted the opposite direction—using the recorded brain activity to predict internal model features—and found the predictions were substantially weaker.
The researchers concluded that AI models can achieve brain‑like output (predicting neuron firing) while relying on computational strategies not employed by the visual cortex. They noted that model components that aligned with brain data were also better at forecasting human behavioral responses in visual tasks.
To enable broader verification, the authors released an open‑source toolkit that implements the reverse predictivity methodology. The toolkit allows other labs to test their own models against neural recordings and to identify which model layers correspond to brain activity.
Dettagli della fonte: journalarta.com
Perché è importante
The findings challenge the common assumption that AI models that match brain activity are using brain‑like computations. By showing that AI can achieve brain‑like predictions through different internal strategies, the study cautions neuroscientists against over‑interpreting model‑brain similarity as evidence of shared mechanisms. This matters for fields that rely on AI to probe visual perception, such as cognitive neuroscience and clinical research into disorders like autism, where model‑based baselines are used. Misreading model similarity could lead to flawed hypotheses about neural processing or ineffective therapeutic targets. The released toolkit gives the community a concrete way to test and validate model‑brain correspondences, promoting more rigorous cross‑disciplinary work.
The study undermines a simplistic view that high predictive alignment automatically implies shared processing mechanisms, a notion that has guided many neuroscience‑AI collaborations.
By exposing this mismatch, the work urges caution in using AI models as direct analogues for brain function, especially when such models inform theories of perception or are used in clinical diagnostics.
The public toolkit promotes transparency and reproducibility, giving researchers a standardized way to assess model‑brain correspondence rather than relying on one‑directional predictions.
Understanding the divergence between model and brain processing could guide the design of next‑generation AI that is both high‑performing and biologically plausible, potentially benefiting explainable AI and brain‑inspired computing.
Meccanismo interattivo: come funziona realmente
Esplora la tecnologia alla base di questo sviluppo in modo interattivo.
Which component of an AI application is the machine-learning model itself?
Cosa guardare dopo
Future work will likely explore which specific model components drive the predictive success and whether alternative architectures can better mirror brain computations. Researchers may apply the reverse predictivity toolkit to other sensory modalities or to clinical populations, testing if model‑brain alignment differs in conditions such as autism or visual impairments. Adoption of the toolkit could also spur development of new AI models explicitly designed to emulate neural processing pathways, potentially improving interpretability and biological relevance.
Adoption of the reverse predictivity toolkit across labs will reveal whether the observed asymmetry holds for other model families (e.g., ‑based vision models) and other brain areas.
Follow‑up studies may investigate whether training models with constraints derived from the toolkit can produce more brain‑like internal representations, narrowing the gap between prediction and mechanism.
Researchers might apply the method to clinical datasets, testing if models trained on neurotypical data fail to predict neural activity in autistic participants, thereby informing diagnostic tools.
The broader AI community may respond by developing architectures explicitly engineered to match known neural computation principles, potentially improving model interpretability and .