We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

We show that high-dimensional neural activity can arise from low-dimensional latent dynamics, both in RNNs and in the brain.

MEIcoder uses most exciting inputs, SSIM loss, and adversarial training to reconstruct high-fidelity visual stimuli from neural population activity, excelling in data- and neuron-scarce settings.