Simple linear models beat complex AI for brain decoding

A new preprint by Ciferri, Ferrante, and Toschi (arXiv:2606.19081) challenges the prevailing assumption that brain decoding—reconstructing stimuli from neural activity—requires complex non-linear models. Drawing on cognitive science theories that represent concepts as high-dimensional vectors, the team tested whether the observable linearity of fMRI signals (due to spatial/temporal averaging) makes simple linear decoders surprisingly effective. They mapped fMRI activity to embeddings from foundation models in vision (e.g., CLIP), language (e.g., BERT), and audio (e.g., Whisper) using contrastive objectives.

The results are striking: linear contrastive decoders consistently outperformed ridge regression and standard non-linear alternatives (e.g., MLPs) across multiple datasets and modalities. The key insight is that decoding performance gains arise from the contrastive training objective—aligning neural representations with model embeddings—not from architectural depth or non-linearity. This suggests that high-dimensional fMRI data, once averaged, behaves like a linear projection of the brain's conceptual space. For AI and neuroscience practitioners, this means simpler, more interpretable models can achieve state-of-the-art brain decoding, potentially accelerating brain-computer interface development and reducing computational overhead. The paper also hints that the same principle may apply to other neuroimaging modalities like MEG or EEG.