How accurate is machine learning in speech recognition? Researchers take a look

Artificial intelligence and machine learning are all the rage—and for good reason. But researchers claim the brain doesn’t actually use the regions identified by machine learning to perform a task. Rather, these algorithms reflect the mental associations related to the task.

One author of the study, published online Jan. 31 in the Proceedings on the National Academy of Sciences, Anne-Lise Giraud with the department of fundamental neuroscience at the University of Geneva in Switzerland, said in a release that "learning algorithms are intelligent but ignorant. They are very sensitive and use all the information in the signals. However, they do not allow us to know whether this information was used to perform the task, or if it reflects the consequences of this task—in other words, spreading information in our brain."

Researchers from the University of Geneva (UNIGE) in Switzerland and the Ecole normale supérieure (ENS) in Paris had 50 people listen to a range of syllables from “BA” to “DA.” Ambiguous options were chosen to make it difficult for participants to distinguish between the options presented.

They used fMRI and magnetoencephalography to visualize how the brain responded when syllables were distinguishable or difficult to differentiate. Researchers noted, despite the ability to differentiate between syllables, a small region of the brain—the posterior superior temporal lobe—was always activated.

To confirm their results, scientists conducted the same analysis on a patient with an injury in the specific region of the posterior superior temporal lobe. He was not able to differentiate between BA and DA.

“[T]his confirms that this small region is important in processing this type of phoneme information,” said corresponding author, Sophia Bouton with the department of fundamental neuroscience at the University of Geneva in Switzerland, in the release.

Maps produced through machine learning suggest the “identity” of the syllable is present in general areas of the brain, while this study demonstrated the information is only present locally. Scientists conducted the same BA-DA study with participants who had electrodes implanted in their brains for medical reasons. The electrodes allowed the neuroscientists to pinpoint neural activity and determine which parts of the brain were at work during the task.

They found their initial conclusions to be correct. When machine learning was applied to the collected data, positive results were indicated in the entire temporal lobe, and beyond it. Not locally as they had discovered.

“The mapped regions outside of the posterior superior temporal lobe are thus false positives, in a way. These regions retain information on the decision that the subject makes (BA or DA), but aren't solicited to perform this task,” said Chambon.