Neuroscience

eLife Assessment

This is a valuable paper that compares various deep learning models, trained with different objective functions, on their ability to predict fMRI data collected during naturalistic video gameplay. The data and analysis provide solid within-distribution evidence that models trained with PPO and imitation learning outperform untrained models and standard convolutional networks. However, the evidence for brittleness in out-of-distribution encoding remains incomplete, as the claim that this stems from the networks' training rather than from alternative causes-like overfitting of ridge regression parameters-is not yet fully supported.

eLife Assessment

In this valuable study, de Vries and colleagues aim to determine how the perception of biological motion is organized at the neural level, specifically testing whether this process rests on hierarchical predictive processing by extending a methodological framework that the authors previously published. The evidence is solid for the empirical claim that neural representations of body motion systematically lead the stimulus in time, with simulations validating the regression approach and consistent effects on both peak magnitude and peak latency. Support for the stronger theoretical interpretation that these signatures specifically reflect active hierarchical predictive inference requires further substantiation, since the design and analysis do not distinguish such inference from cached associative retrieval or from nonlinear temporal integration of slowly varying features.