For years, headlines said that Japanese scientists built an MRI dream machine. The research is real, but the claim is not. Two lines of work from Japan explain what is possible today.
First, in 2013, a team in Kyoto used functional MRI, a brain scan that tracks blood flow as a stand-in for neural activity, to predict broad elements of what people were dreaming. The system matched brain patterns during light sleep with patterns seen when the same people were awake and looking at pictures. It could tell, with some success, whether a dream likely included simple objects like a person, a car, or a house. It did not create pictures of dreams, and it did not read thoughts.
Second, in 2023, an Osaka group showed that images seen while awake can be roughly rebuilt from MRI data with help from a modern image generator called a latent diffusion model, known as Stable Diffusion. Their method links patterns in the visual cortex to the internal features of the AI model, so the AI can produce a picture that reflects the broad shapes and meanings of what the person viewed. This is not mind reading, and the study did not test dreams. The authors state clearly that it cannot read private thoughts and that use with dreams or pure imagination is still unclear.
Taken together, these studies show steady progress in decoding limited visual content from the brain, under strict lab conditions, for a trained individual, inside a large MRI scanner, for short and simple cases. They do not add up to a machine that records dreams like a movie, and they do not work across people without retraining.
Neural decoding of visual imagery during sleep – 2013
Using MRI during light sleep, the team trained machine-learning models on each person’s brain activity while awake, then predicted broad categories likely present in the person’s dream reports. No images were generated, and decoding was limited to simple content.
Neural Decoding of Visual Imagery During Sleep, project page – 2013
Lab overview of the Kyoto dream-decoding study, explaining how decoders trained on waking vision were applied to sleep data to infer coarse dream categories.
High-resolution image reconstruction with latent diffusion models from human brain activity – 2023
Shows that AI image generators can rebuild approximate images viewed while awake from MRI signals by mapping brain activity to the model’s internal features. The work does not decode dreams.
0 Comments