In the realm of artificial intelligence, we have been exploring how to make machines think more like humans. Now, researchers from the Georgia Institute of Technology have taken a significant step forward by developing the first neural network capable of simulating human perception and decision-making processes—the RTNet.
The birth of RTNet marks an important advancement in our understanding and simulation of how the human brain works. This new neural network can not only generate random decisions but also mimic the human response time distribution, something that was previously unseen in AI models.
Unlike previous neural networks, RTNet adjusts its "thinking" time based on the difficulty of the task. Just as we can quickly provide answers to simple questions but need more time to ponder complex ones.
Image Source: The image was generated by AI, and the image authorization service provider is Midjourney.
The internal mechanism of RTNet includes two stages. The first stage uses the Alexnet architecture, but the weight parameters exist in the form of a Bayesian Neural Network (BNN), introducing randomness. The second stage is an accumulation process, where reasoning stops once a certain class reaches the threshold.
Comprehensive tests have shown that RTNet can replicate all the basic characteristics of humans in terms of accuracy, response time, and confidence, and it outperforms other existing models.
In experiments, 60 participants performed digit discrimination tasks and assessed their decision-making confidence. Meanwhile, RTNet was compared with several advanced neural networks.
The experimental results indicate that RTNet excels in simulating the randomness of human decision-making and can adjust its response time based on task difficulty. In contrast, the decision-making processes of other neural networks are entirely deterministic.
The success of RTNet is not only technically significant but also provides a new perspective on understanding the workings of the human brain. Its concept is similar to the race model in cognitive models but has advantages in image computability and capturing relationships between choices.