Decoding Papers

A selection of brain decoding papers spanning language, images, video, and speech — across fMRI, MEG/EEG, ECoG, and intracortical recordings. Columns capture modality, what was decoded, dataset scale, model, and reported metrics.

PaperModalityWhat decoded# participantsData (hours)ModelMetrics
Semantic reconstruction of continuous language from non-invasive brain recordingsfMRIcontinuous language (semantic gist of stories)3~16 h / subjectencoding model + LM + beam searchsemantic similarity, timepoint accuracy (~65–82%)
Generative language reconstruction from brain recordingsfMRIcontinuous language (direct generation)5 + 8 + 28 (three datasets)multi-hour per subjectLLaMA-2 conditioned on brain signalsBLEU / ROUGE / semantic similarity (paper-dependent)
Reconstructing the Mind's Eye: fMRI-to-Image with Contrastive Learning and Diffusion PriorsfMRIimages (natural scenes)4 (NSD subjects)~40 sessions / subjectCLIP + diffusion prior>93% retrieval top-1, CLIP similarity
High-resolution image reconstruction with latent diffusion models from human brain activityfMRIimages4~30–40 sessionsStable Diffusion latent regression~75–83% identification accuracy
Brain decoding: toward real-time reconstruction of visual perceptionMEGimages4 (THINGS-MEG)~4 sessions / subjectcontrastive MEG encoder + latent diffusion7× retrieval improvement over linear, perceptual similarity
Reconstructing visual experiences from brain activity evoked by natural moviesfMRIvideo clips3multi-sessionmotion-energy + Bayesian prior~95% identification
Deep image reconstruction from human brain activityfMRIimages + imagery3months (~2h/day sessions)VGG feature inversionhuman eval ~80%+
Neuroprosthesis for decoding speech in a paralyzed personECoGspeech → text1~22 hneural classifier + LM~15 wpm, ~25% WER
Generalizable spelling using a speech neuroprosthesisECoGspelling (text)1multi-sessiondeep net + LM~6% CER
A high-performance speech neuroprosthesisintracorticalspeech → text1multi-sessionRNN decoder + LM rescoring9.1% WER (50 vocab), 23.8% (125k)
Decoding speech perception from non-invasive brain recordingsMEG/EEGperceived speech segments~175large aggregated datasetscontrastive + wav2vec44% top-1 MEG / 19% top-1 EEG (1,594 segments)
Brain-to-Text Decoding: A Non-invasive Approach via TypingMEG/EEGtyped sentences35multi-sessionCNN + transformer + character LM19–32% CER (MEG), 67% CER (EEG)