Dozens of scientific papers have been debunked because of blind spots discovered at the intersection of AI and neuroscience.

 Is it possible to read someone's mind by decoding their brain's electrical signals? The response may be much more complicated than most people believe.

Purdue University researchers working at the intersection of artificial intelligence and neuroscience claim that a well-known dataset used to try to address this question is skewed, and that many eye-popping results based on it that have gained high-profile attention are, in fact, inaccurate.

The Purdue team ran comprehensive experiments on the dataset for more than a year, looking at the brain function of people who took part in a study where they looked at a series of photographs. While viewing the photographs, each individual wore a cap with hundreds of electrodes.

“This measuring technique, known as electroencephalography or EEG, may provide knowledge about brain activity that could, in theory, be used to read minds,” said Jeffrey Mark Siskind, a Purdue College of Engineering professor of electrical and computer engineering.

“The issue is that they used EEG in such a way that the dataset was tainted in the first place. The researchers were able to tell what image was being seen only by reading the timing and order information found in EEG, rather than solving the real problem of interpreting visual perception from brain waves, since the analysis was performed without randomising the order of images.

When the Purdue researchers couldn't replicate the results in their own experiments, they started challenging the dataset. That's when they began looking at the previous findings and discovered that the dataset had been corrupted due to a lack of randomization.

Hari Bharadwaj, an assistant professor with a joint appointment in Purdue's College of Engineering and College of Health and Human Sciences, said, "This is one of the challenges of working in cross-disciplinary research areas." “Important scientific questions often necessitate collaboration across disciplines. The problem is that researchers who have been educated in one area might not be aware of the common pitfalls that can arise when their ideas are applied to another. In this case, the previous work seems to have suffered from a mismatch between AI/machine-learning scientists, as well as pitfalls that neuroscientists are well-aware of.”

The Purdue team looked at papers that used the dataset for tasks like object classification, transfer learning, and image production based on brain-derived representations measured by electroencephalograms (EEGs).