Scaling Parameter-Efficient Fine-Tuning Boosts Brain Encoding
fine-tuning training
| Source: HF Papers | Original article
Researchers propose a parameter‑efficient fine‑tuning method to scale brain‑encoding models, tackling the difficulty of training large models with limited neural data.
A new study demonstrates that parameter‑efficient fine‑tuning (PEFT) can unlock the predictive power of massive multimodal models for brain‑encoding tasks. Researchers from Columbia, Mind Lab, Tsinghua and HKUST(GZ) show that, despite the high cost and scarcity of fMRI data, PEFT enables effective adaptation of backbones as large as 104 billion parameters. By applying a causal multi‑timepoint extraction technique, the team achieves state‑of‑the‑art performance on the Algonauts 2025 challenge and makes it feasible to run ensembles of up to 200 models.
The finding matters because the prevailing strategy in neural‑prediction work—freezing a pretrained backbone and training only a readout layer—leaves much of the model’s capacity untapped. Fully fine‑tuning such large models has been impractical given the limited neural datasets available. PEFT bridges this gap, allowing researchers to leverage the rich representations of today’s biggest language‑vision models without overwhelming the data budget. The resulting boost in accuracy could sharpen our understanding of how the brain processes complex stimuli and accelerate the development of brain‑computer interfaces that rely on precise neural decoding.
Going forward, the community will watch for broader validation of the approach across other neuroimaging benchmarks and modalities. Open‑source releases of the PEFT pipelines could spur rapid adoption, while extensions to real‑time decoding or clinical applications may test the limits of large‑scale brain‑encoding models. The paper signals a shift toward more data‑efficient exploitation of ever‑larger AI backbones in neuroscience.
Sources
Back to AIPULSEN