Kaiyue Pang

V model, yielding final reconstruction. Motion Adaptation Decoder (MAD) To improve the fidelity of motion dynamics reconstructed from brain signals, we propose a semantically guided motion adaptation decoding strategy. Let the original video stimulus be denoted as V={v1,v2,⋯,vn}V=\{v_{1},v_{2},\cdots,v_{n}\}, where nn is the number of frames. We extract its motion latent representation using a pretrained VAE encoder 𝐄\mathbf{E}, yielding E​(V)={e1,e2,⋯,en}E(V)=\{e_{1},e_{2},\cdots,e_{n}\}. The objective of MAD is to take the fMRI signal XX as input and generate a sequence of predicted motion latents corresponding to the reconstructed video frames, denoted as E^​(X)={e^1,e^2,⋯,e^n}\hat{E}(X)=\{\hat{e}_{1},\hat{e}_{2},\cdots,\hat{e}_{n}\}. To achieve this, MAD first projects the fMRI signal into the latent space using a subject-specific projection network fMADprojf_{\text{MAD}}^{\textit{proj}}, implemented as a multilayer perceptron