Jakarta, INTI - Researchers from Stanford University in the United States and Tel Aviv University (TAU) have developed an artificial intelligence (AI) training method that enables computer systems to improve their predictive capabilities without relying on large collections of real-world data.
The findings, published on the scientific preprint platform arXiv, introduce a new approach to developing machine learning capabilities.
In the study, scientists trained two computer models simultaneously using sequences of symbols as learning material.
The results showed that the models' predictive performance improved as the computing power allocated to the training process increased.
Two AI Models Trained Simultaneously Using Symbol Sequences
The researchers implemented a training method involving two computer models with different but interconnected tasks.
The first model was responsible for generating sequences of symbols that served as training material for the second model.
Meanwhile, the second model was trained to identify patterns within these symbol sequences and predict the next symbol.
Throughout the training process, the first model gradually improved its ability to generate more effective learning materials.
This development enabled the second model to strengthen its predictive capabilities through training data generated by another computer system.
The approach allows the learning process to take place without relying primarily on large collections of human-generated training materials.
AI Capabilities Tested Across Language, Images, and DNA
Following the training process, the research team evaluated the AI models using several types of information to assess the effectiveness of the proposed method.
The tests covered language, images, audio, music, computer code, and DNA sequences.
The results indicated that the models' predictive capabilities continued to improve as the computational resources allocated during training increased.
These findings suggest that training with symbol sequences can help AI systems develop predictive capabilities applicable to various types of information.
However, the results do not mean that AI systems can understand all aspects of the real world without learning from information related to it.
Researchers Emphasize AI Still Needs Real-World Data
The researchers highlighted several limitations of the training method, particularly when it comes to learning facts and knowledge about the real world.
AI systems trained exclusively on symbol sequences cannot acquire factual knowledge about the world without exposure to relevant information.
Therefore, the approach is not intended to completely replace real-world data in AI development.
Instead, the method could potentially help AI systems learn new information more efficiently when they are subsequently exposed to real-world data.
The study published on arXiv opens opportunities for further research into AI training methods that do not entirely depend on massive collections of human-generated data.
The findings also offer insights into the potential development of more efficient machine learning systems through synthetic training data and increased computational power.
Conclusion
Researchers from Stanford University and Tel Aviv University have demonstrated a new AI training approach using synthetic symbol sequences, with predictive capabilities improving alongside increased computing power. While real-world data remains essential for factual understanding, the findings could pave the way for more efficient AI training methods.
Read more: Komdigi Urges Local Governments to Leverage AI for Better Productivity and Public Services