Groundbreaking AI “reads your mind”: Turns thoughts into images with 75% accuracy

In a groundbreaking study, researchers in Japan have reconstructed images from human brain activity using artificial intelligence. While this isn’t the first study of this kind, it is the first that managed to reconstruct images without the subject looking at them. In other words – only their thought was enough, and AI achieved an astonishing accuracy of over 75%!

In previous studies, image recreation from brain activity was only possible in two scenarios. One, when subjects were directly viewing the images. And two, when the type of images was predetermined (for example, faces, letters, or simple figures). But the team of researchers from the National Institutes for Quantum Science and Technology (QST) is the first one to overcome these boundaries. Their model reconstructed a wide array of images, including complex landscapes and figures, based solely on thought.

The process: From brain activity to AI image reconstruction

The researchers trained an algorithm with two input sets based on the process reported in The Mainichi. The scientists first had their subjects view 1,200 various images while inside a functional magnetic resonance imaging (fMRI) machine. The AI then created “score charts” from their brain activity, encompassing 6.13 million factors like color, shape, and texture. A neural signal translator program then matched brain activity with these scoring charts, updating them with new brain activity inputs.

The next step was to show the subjects a different image from the original 1,200. The scientists asked them to imagine it while their brain activity was measured. The neural signal translator used these records to create new score charts. After that, they fed them into another generative AI program.

Finally, this program underwent a 500-step revision process to reconstruct the image. As I mentioned, it achieved a remarkable 75.6% accuracy rate in identifying original images from the reconstructed ones. We no longer need to read Sci-Fi novels, we’re living them.

Previous studies

The Netherlands, 2022: At Radboud University, researchers developed a technology that turns brainwaves into photographic images with remarkable accuracy. Thirza Dado, an AI and neurocognitive scientist, led the project. The process involved fMRI scans of brain activity while exposed to visual stimuli (e.g., photos of faces). The scientists then trained an AI algorithm with this data to reconstruct visual images. The results, published in Nature magazine, showed impressive stimulus reconstructions and hinted at various future applications.

Japan, early 2023: Another Japanese team from Osaka University introduced an AI tool that converts brain activity into high-resolution images without the need for training or fine-tuning the AI models. Their method, detailed in a paper published in December 2022, relied on a latent diffusion model called “Stable Diffusion” to reconstruct images from fMRI data. This method showcased eerily accurate results and provided a new framework for understanding diffusion models in AI.

Implications and future prospects

This research opens up many possibilities, and I don’t know where to start. We can develop new forms of wordless communication, which could be a game-changer for those with any speech difficulties. It could also help us better understand the human mind, including our dreams (yes, I still want that “dream machine” I imagined as a child).

“This is a monumental achievement in which humans peered into another person’s head for the first time,” said QST researcher Kei Majima. “I hope this research prompts further understanding of the human mind.” The study, titled Mental Image Reconstruction from Human Brain Activity: Neural Decoding of Mental Imagery via Deep Neural Network-Based Bayesian Estimation, was published in the international science journal Neural Networks.

[via PetaPixel; lead image was AI-generated]

---