The human brain has long been considered a mess of biological signals that humans have no hope of untangling. However, a study published in the Oct. 29 issue of Cerebral Cortex by the Carnegie Mellon psychology department indicates that the human brain’s representation of abstract concepts is not only identifiable, but also shared.
The study was one of several by the Center for Cognitive Brain Imaging (CCBI). It used machine learning to analyze the functional magnetic resonance image (fMRI) scans of nine individuals while they were instructed to think about 28 abstract concepts, such as subtraction, causality, ethics, gravity and gossip. While past studies by CCBI have tended to focus on concrete concepts, the researchers’ refinements have allowed them to conclude that abstract concepts are also identifiable across participants.
“When fMRI came about, first people just studied where the activation was in the brain while you were doing something.… The big breakthrough occurred in some work done at Carnegie Mellon with my colleague Tom Mitchell in the machine learning [department], where we found not only where the hotspots are, but… what the hotspots were coding. So we found out how the concept of ‘apple’ was represented. Once we had that, the sky was the limit,” said Marcel Just in an interview with The Tartan. Just is the senior author of the study, professor of Psychology at the Dietrich College of Humanities of Social Sciences, and the director of CCBI. “The goal is to develop an ontology of human knowledge that is neurally based,” Just said. “It’s not just a big mess. It’s in there systematically as it is in a library. Only we know the library’s indexing system: you just have to go to the card catalogs. So the goal is to learn the brain’s card catalog… Each of our studies gives us another drawer in the card catalog, and this study tells us how abstract concepts are organized.”
“We want to know how the brain works, how knowledge is represented. We think we’re at the point now where we can start to identify knowledge representations from the conceptual level … the next stage is combinatorics: can we see how conceptual representations are combined while solving problems,” said Robert Mason, a senior research psychologist at CCBI who has worked with Just on related studies.
The researchers used Gaussian Naïve Bayes, a conventional machine learning algorithm for classification, to compare the MRI scans of each individual. Each individual was told to think about each of the 28 concepts on six separate occasions. Four of the six scans were used to train the classifier algorithm, while the average of the other two was used to test its validity. The same procedure was used across participants, using eight of the nine to train the algorithm, and the ninth to test accuracy. The process was repeated with different combinations of scans in the training set in order to confirm the accuracy of the algorithm.
Traditionally, machine learning requires inordinately large data sets. The study only used nine participants, each of which had six rounds of scans. The way the researchers got around this was essentially by using each pixel of the MRI scan as a classifier and ranking all the pixels based on how strong a correlation they had to the result, then picking the most correlated ones as features of the algorithm.
There are many potential uses for this use of fMRI technology, ranging from education to therapy. Just explained, “Psychiatry is the study of the disorder of thought, and here we are claiming we can measure thoughts. Well, if we can measure thoughts, we can measure the disorder of thoughts. So we have a large ongoing project where we identify people who have been thinking about suicide by looking at their neural representation of certain concepts and seeing if they are different from controls. So we can identify whether someone is suicidal or not with 91 percent accuracy.”
“We really think it’s possible to use these fMRI images as an assessment tool,” Mason added. “So, for example, we can see how velocity is represented at the beginning and end of the semester. We can compare those images to students who got A’s, or to professors, and we can see if the brain images are similar to the experts. But the real place that it could have some type of impact, years from now, is for assessing learning readiness and tutoring. So if we have a student who is struggling, maybe we can identify where their knowledge base is failing.”
Currently, Mason is working on a study with nearly identical tactics to Just’s but focused on physics concepts, attempting to test the validity of the educational approach. According to Mason, the study will compare scans of students in first-year physics from the beginning to the end of the semester “to see if there is a change in the representations that is indicative of learning.”
Another study is in progress with a similar scope as Mason’s, except that the participants are physics professors, and the words include advanced physics concepts that are almost impossible to conceptualize concretely.
"brain" - Google News
November 10, 2019 at 05:26PM
https://ift.tt/2K9yGNC
Machine learning unlocks library of the human brain - CMU The Tartan Online
"brain" - Google News
https://ift.tt/2W2PMS9
Shoes Man Tutorial
Pos News Update
Meme Update
Korean Entertainment News
Japan News Update
No comments:
Post a Comment