Scientists have developed a newAItraining methods that provide AI with a basic human skill: connecting two learned concepts into a new one. Their combinatorial meta-learning (MLC) allows the AI to repeat combinatorial tasks. Soon, it may eliminate the need to retrain the machine every time it encounters a new concept.
Artificial intelligence companies like OpenAI recognize that their work will eventually lead to a machine that can think and act like a human. Their research provides the world with generative AI tools that can produce almost any media we want. Soon, AI may be able to generate new ideas without human intervention, paving the way for even more opportunities.
In this article, I will discuss this revolutionary AI training technique. Later, I'll explain how the AI training technique has been used since ChatGPT How much progress has been made in AI since it became mainstream.
How are scientists creating new AI training?
Before addressing MLC, let's briefly discuss how generative AI tools work.ChatGPT and similar tools rely on large language models containing large numbers of words.
It matches words with the user's words and combines them into coherent, relevant answers. For example, if it receives the word "jump", it can make phrases such as "jump twice" or "jump right twice".
However, if the AI program gets an unknown word, such as "spin", it will not provide results. You have to retrain the entire LLM for that word, which is a painstaking and expensive process.
That's why scientists have created a new way to train AI: combinatorial meta-learning. As mentioned earlier, it makes AI tools apply different rules to newly learned words.
It also provides feedback on whether it followed the rules correctly. The researchers used the following steps to test their AI training methods:
They have humans match the same words using the same rules.
Then they documented human error.
After that, they instructed the AI to learn as it completed the task. Traditional methods involve tracking static datasets.
Experts compare AI and human performance by applying human error to their AI.
The answers shared by the AI program were almost identical to those of the human volunteers.
"For 35 years, researchers in cognitive science, artificial intelligence, linguistics and philosophy have been debating whether neural networks can generalize human-like systems," said NYU scientist Brenden Lake.
"We show for the first time that general-purpose neural networks can mimic or exceed human system generalization in head-to-head comparisons.
Ella Bruni, an expert in natural language processing at the University of Osnabrück in Germany, said the research could make AI programs more efficient learners.
What are the capabilities of modern AI?
ChatGPT has shown the world what artificial intelligence is capable of. It surprised everyone and produced almost any text imaginable, from jokes to research papers....
However, it has become so advanced that scientists are struggling to filter artificial research papers from AI-generated papers. Today, more and more experts are going through AI-generated research, which poses new challenges for peer review.
Dr. Catherine Gao, a physician and scientist at Northwestern University, tested whether ChatGPT could create compelling research papers. "I wanted to know if it could write scientific abstracts," Gao said.
"I asked it to write an abstract about a hypothetical machine learning study focusing on pneumonia in the ICU," she adds. As a result, the program surprised her with a "very good abstract."
ChatGPT showed surprisingly high emotional awareness prior to the Meta-learning for Compositionality study. At the time of writing, the AI program was unable to display or report emotions.
Zohar Elyoseph, Dorit Hadar-Shoval, Kfir Asraf, and Maya Lvovsky presented scenarios of the Emotional Awareness Level Scale.
It usually involves human respondents imagining themselves in various scenarios and writing down their "you" emotions. Artificial intelligence researchers have replaced "you" with "human" because this approach is not applicable to machine learning models.
Separate test sessions are available to help experts validate the results. The first generated Z-score was 2.84 and the second was 4.26. A Z-score is a statistical measure of how close a value is to the median or mean score.
A Z score above 1 indicates a higher than average value, meaning that ChatGPT demonstrates a higher emotional awareness than most people. In addition, it answered more accurately than humans, earning a score of 9.7 out of 10.
reach a verdict
Researchers have created a new AI training method that provides AI with a crucial human skill. It enables AI tools to incorporate learned concepts into new concepts.
Previously, we had to retrain AI models every time a human invented a new idea or object. Soon, AI can create new concepts on its own.