The practice of talking to oneself, often perceived as a human characteristic, could become a major asset for artificial intelligences. Just as this internal dialogue helps us reflect or choose a direction, it also allows AI systems to learn more effectively and adjust to new circumstances with a reduced amount of information.
A study published in
Neural Computation demonstrates that combining this internal monologue with a specially designed short-term memory significantly improves model performance.
The working memory and inner speech architecture improves AI performance on complex pattern generation challenges.
Credit: Kaori Serakaki/OIST
Scientists at the Okinawa Institute of Science and Technology observed that this method allows systems to process multiple tasks in parallel and solve complex problems more easily. To achieve this, the team organized the training data to teach the AI the art of talking to itself.
This approach relies on a working memory that temporarily holds data, similar to our brain when following instructions or performing quick calculations. Tests compared different memory structures, revealing notable differences. Models equipped with multiple temporary memory slots performed better on certain challenges, such as reversing sequences or reproducing patterns. They were able to keep multiple items "in mind" and manipulate them.
The introduction of internal 'mumbling' objectives, where the system is prompted to talk to itself a defined number of times, further increased its efficiency. The most significant progress was recorded in multitasking and multi-step problems. Furthermore, this combination works even with limited datasets, unlike the large sets usually required for training. It thus offers a lightweight and complementary alternative.
The researchers now plan to apply this method to more realistic and less structured environments. Indeed, choices are often made in noisy and unpredictable settings. Reproducing these conditions would bring it closer to human developmental learning. This breakthrough also helps clarify certain brain mechanisms, paving the way for concrete applications.