MIT Artificial Intelligence Program Has IQ Of 4-Year-Old Child
9 October 2015, 3:27 am EDT
By Katherine Derla Tech Times
The Massachusetts Institute of Technology’s artificial intelligence program has a machine called the ConceptNet. Researchers from the University of Illinois put it to the test and found its IQ is the same as a 4-year-old’s. ( Rudy And Peter Skitterians | Pixabay )
Machines with artificial intelligence have shown quite an improvement in several aspects. For instance, they can play chess, compute complex mathematical problems and even identify pictures.
The Massachusetts Institute of Technology (MIT) has created an intelligent machine called ConceptNet under its artificial intelligence program. Researchers from the University of Illinois did a study on the ConceptNet 4 version of the AI and they discovered that its IQ level is the same as that of a 4-year-old child.
The research team, led by Stellan Ohlsson, chose an IQ test known as the Wechsler Preschool and Primary Scale of Intelligence. The IQ test is commonly used in schools across the United States and is designed to measure the level of intelligence in five key categories.
The first one is information, in which participants need to answer questions such as "Where can you find a giraffe?" The second category is vocabulary and includes questions such as "What is a giraffe?" The third category is word reasoning wherein participants are given three clues about a word they need to define. The fourth category is similarities where participants need to answer questions such as "Giraffes and zebras are both...?" The final category is comprehension where participants need to answer questions in the form of an explanation. In this example, participants can be asked, "Why do giraffes run when they see a lion?"
The same IQ test was given to ConceptNet 4; however, several modifications were done for the machine to understand the questions.
Researchers found that ConceptNet 4 handled vocabulary and similarities well. In the information category, ConceptNet4 got an average score, while it got poor scores in comprehension and reasoning.
"The ConceptNet scored a [mark] that is average for a 4-year-old child, but below average for 5- to 7-year-olds," concluded the researchers. The result was highly dependent on how the intelligent machine understood and interpreted the questions.
As an example, ConceptNet 4 was asked why people shook hands. The machine gave the answer "epileptic fit." The researchers revised the question and reduced it to two words: "shake hands." ConceptNet answered with more relevant words such as "meet friend," "thanks" and "flirt."
Some of ConceptNet 4's responses appear to be unrelated. For instance, it was asked "Where can you find a teacher?" The machine answered "band" and "piano."
While the researchers could not put a finger on the irregularities in ConceptNet 4's answers, they suggested that a Siri-like virtual assistant can help the machine to understand the questions better using its natural language.
"The ConceptNet system scored a WPPSI-III VIQ that is average for a 4-year-old child, but below average for 5- to 7-year-olds. Large variations among subtests indicate potential areas of improvement," the researchers wrote in the paper's introduction.
The findings were also published in the MIT Technology Review on Oct. 1.
By Katherine Derla Tech Times
Machines with artificial intelligence have shown quite an improvement in several aspects. For instance, they can play chess, compute complex mathematical problems and even identify pictures.
The Massachusetts Institute of Technology (MIT) has created an intelligent machine called ConceptNet under its artificial intelligence program. Researchers from the University of Illinois did a study on the ConceptNet 4 version of the AI and they discovered that its IQ level is the same as that of a 4-year-old child.
The research team, led by Stellan Ohlsson, chose an IQ test known as the Wechsler Preschool and Primary Scale of Intelligence. The IQ test is commonly used in schools across the United States and is designed to measure the level of intelligence in five key categories.
The first one is information, in which participants need to answer questions such as "Where can you find a giraffe?" The second category is vocabulary and includes questions such as "What is a giraffe?" The third category is word reasoning wherein participants are given three clues about a word they need to define. The fourth category is similarities where participants need to answer questions such as "Giraffes and zebras are both...?" The final category is comprehension where participants need to answer questions in the form of an explanation. In this example, participants can be asked, "Why do giraffes run when they see a lion?"
The same IQ test was given to ConceptNet 4; however, several modifications were done for the machine to understand the questions.
Researchers found that ConceptNet 4 handled vocabulary and similarities well. In the information category, ConceptNet4 got an average score, while it got poor scores in comprehension and reasoning.
"The ConceptNet scored a [mark] that is average for a 4-year-old child, but below average for 5- to 7-year-olds," concluded the researchers. The result was highly dependent on how the intelligent machine understood and interpreted the questions.
As an example, ConceptNet 4 was asked why people shook hands. The machine gave the answer "epileptic fit." The researchers revised the question and reduced it to two words: "shake hands." ConceptNet answered with more relevant words such as "meet friend," "thanks" and "flirt."
Some of ConceptNet 4's responses appear to be unrelated. For instance, it was asked "Where can you find a teacher?" The machine answered "band" and "piano."
While the researchers could not put a finger on the irregularities in ConceptNet 4's answers, they suggested that a Siri-like virtual assistant can help the machine to understand the questions better using its natural language.
"The ConceptNet system scored a WPPSI-III VIQ that is average for a 4-year-old child, but below average for 5- to 7-year-olds. Large variations among subtests indicate potential areas of improvement," the researchers wrote in the paper's introduction.
The findings were also published in the MIT Technology Review on Oct. 1.
