Researchers at Open AI have developed an AI system capable of tackling elementary math problems (the company behind the GPT-3 language model). This is a big step forward because until now, models could only use language to generate words, not the multi-step thinking required to solve math problems. This is now completed.
What is Data Science and Artificial Intelligence
The two most important technologies in the world now are data science and Artificial Intelligence. While Data Science employs AI in its processes, it does not fully represent AI.
Methods and Goals in AI
Connectionist vs. symbolic approaches
The symbolic (or “top-down”) approach and the connectionist (or “bottom-up”) approach are two distinct, and to some extent competing, approaches to AI research. The top-down approach aims to reproduce intelligence by examining cognition in terms of symbol processing, regardless of the organic structure of the brain—hence the symbolic moniker.
The bottom-up technique, on the other hand, entails the creation of artificial neural networks that mimic the structure of the brain—hence the connectionist moniker.
Consider the task of developing a system with an optical scanner that recognizes the letters of the alphabet to demonstrate the differences between these approaches. A bottom-up strategy involves gradually improving performance by “tuning” an artificial neural network by presenting letters to it one by one. (Tuning changes how receptive different brain pathways are to certain stimuli.)
According to Researchers on AI
Researchers have trained a model to detect its own errors, allowing it to re-evaluate its responses until it finds a satisfactory solution. In tests, the AI system was able to solve nearly as many issues as a group of nine to twelve-year-old youngsters. The children received a 60 percent on an Open AI database test, while the AI system received a 55 percent.
Although elementary school math is very simple, the emergence of AI models that can handle fundamental math issues is a significant step forward for Open AI and opens up numerous possibilities. According to the researchers,
“The high sensitivity to individual errors is a key difficulty in mathematical thinking.” Autoregressive models have no method for correcting errors because they create each solution token by token. When things go wrong, it’s difficult to recover.”
This challenge was solved by introducing checkers whose job it was to evaluate the AI’s responses. These checkers were given a list of 100 possible answers generated by the model and asked to determine which ones were correct. The researchers from Open AI add:
“In the creation of more broad AI, providing right arguments and recognizing wrong ones are significant issues. Although the problems are essentially simple, one small error might derail a whole solution. Detecting and eliminating such errors is a critical skill for model building.”
Open AI believes that as AI is applied to more complicated fields, the verification system that allows its AI systems to solve simple mathematical problems with some accuracy will become more effective.
It will be conceivable to imagine AI models that are considerably larger than they are today if this study is combined with developments in semiconductors. As a result, the power to adjust how AI approaches a problem could be game-changing. Now masters in data science includes more of such research for you to explore.
Leave a Reply