A group of scientists conducted an experiment to study the logic and memory of artificial intelligence.
The research showed that in neural networks, memory and analytical abilities function separately. Even after the removal of certain knowledge, the AI demonstrated similar accuracy in reasoning.
The team from the startup Goodfire.ai decided to investigate how AI thinking works by disabling part of its memory and examining whether its ability for logical conclusions would change.
For this experiment, the scientists applied the K-FAC (Kronecker-Factored Approximate Curvature) mathematical method, which allows for selective data removal without disrupting the overall structure of the neural network.
After the "amnesia" procedure, the AI could still name the capitals of countries, but made mistakes in the names of company leaders, which reduced the accuracy of its answers in this category by 78%.
At the same time, the AI's abilities to solve mathematical and logical problems changed little, confirming the possibility of separating memory and reasoning.
The scientists also noted that the more frequently the model encountered information during training, the better it remembered it—rare facts were erased more quickly.
The researchers are convinced that the results obtained will aid in the creation of AI that can be partially "reset," removing unnecessary data without affecting its analytical capabilities.
