Artificial intelligence researchers have long dreamed of creating knowledgeable and communicative computers like the Star Trek computers. It can interact with humans in natural (i.e., human) language. Last week we seemed brave to move towards that goal. The New York Times reports that a team from the Allen Institute for Artificial Intelligence (AI2) has reached a “milestone in the field of artificial intelligence.” Aristo’s AI2 program not only passed but also administered a standardized science test for eighth graders. The machine, the Times reported, was “ready for high school science. Maybe even college. This post will discuss how the AI system helps computers ace different tests help IQ.
Or maybe not. Aristo is not the first AI system to shine on tests designed to measure human knowledge and thinking abilities. In 2015, a system performed as well as a four-year-old on an IQ test, leading to the BBC headline ‘AI has the IQ of a four-year-old.’ Another group reported that their system could solve SAT geometry questions “as well as the average American eleventh grader.” Recently, Stanford researchers created a question-and-answer test that led the New York Post to state that “artificial intelligence systems outperform humans in reading comprehension.” The truth is that while these systems perform well on specific language processing tests, they can only pass the test. Nothing comes close to reading comprehension or other general skills the tests are design to measure.
The Working of Artificial Intelligence in Machines
The problem is that today’s machines, which excel at specific rigorous tasks, still lack common sense. It includes the extensive and largely unconscious background knowledge we use to understand the situations we find ourselves in and the language we use to communicate. Common sense also has our ability to apply this knowledge quickly and flexibly to new circumstances.
The goal of giving engines common sense is as old as the field of AI itself and, I dare say, the most complex AI problem to spot. Starting in the 1990s, common sense research gave way to data-driven statistical approaches to AI—especially in the form of neural networks and deep learning. However, researchers have recently found that deep learning systems lack the resilience and generalizability of human understanding, primarily because they lack broad knowledge and flexible thinking skills. Incorporating human-intelligent machines is now a top priority on the AI to-do list.
Artificial Intelligence Examples
Answering open-ended questions like the Star Trek computer is still too difficult for today’s artificial intelligence systems. Therefore, researchers are making progress by creating programs that perform well on “benchmarks” — specific data sets representing a particular task. The Aristo Benchmark consists of multiple-choice questions from the New York State Regents Science Exam. It also answers that computers ace different tests to help IQ.
Aristo’s creators believe that developing an AI system to answer such questions. It is one of the best ways to advance this field. “While not a complete machine intelligence test,” they note, “these questions test several skills closely related to intelligence, including language comprehension, reasoning, and use of common sense.”
Aristo (Artificial Intelligence System)
Aristo is a complex system that combines several AI methods. However, the component responsible for nearly all of the system’s success is a deep neural network trained to become a language model. This mechanism can use a word order to predict the next word. “I was speeding when I stopped…” What was the next term? Maybe “police.” Perhaps not “grapefruit.” Given the order of words, the language model calculates the probability that each of the hundreds of thousands of words in its vocabulary is the following sequence.
Neural networks are notoriously opaque; knowing precisely what they have learned is usually challenging. It could be that Aristotle’s brilliant achievement was actually due to his ability to derive and justify scientific concepts. But given the history of natural language processing systems exploiting bounties and being “true for the wrong reasons.” It’s essential to examine these claims more thoroughly. Aristo’s team offered an open step in this direction: they ran an experiment in which they added four additional incorrect answers to each question, specifically selecting new solutions that might confuse the system. Aristo’s performance dropped to less than 60 percent correct. Learning the weaknesses of one’s AI system is essential to progress in this complicated matter. We hope now you know computers ace different tests help IQ.
