People make thousands of decisions every day, from simple ones, like crossing the street, to more complex ones, such as food choices. Researchers at Georgia Tech have developed a neural network that mimics human decisions, aiming to get closer to the human way of thinking and decision-making. This network, called RTNet, uses a Bayesian neural network (BNN) and an evidence accumulation process to make decisions in a way similar to humans.

Decoding Decisions
"Neural networks typically make decisions without expressing their level of confidence in those decisions," said Farshad Rafiei, a psychology doctor from Georgia Tech. "This is one of the key differences compared to humans." However, this new network can provide answers that include a degree of confidence, which is a crucial step toward human-like decision-making behavior.

Large language models (LLMs) often fabricate answers when they do not know the correct information. Unlike them, humans will admit ignorance in similar situations. Building networks that better mimic human reactions can reduce this type of error and improve the accuracy of responses.

Model Building
The team at Georgia Tech trained their network on handwritten digits from the well-known MNIST dataset. To test the accuracy of the model, they added noise to the images, making it harder to recognize the numbers. The model was then compared to the results of human subjects. Sixty students observed the same images and expressed their confidence in the decisions, and the results showed similarities in accuracy, reaction time, and confidence patterns between humans and the network.

The researchers used two key components: BNN, which uses probability for decision-making, and the evidence accumulation process, which tracks evidence for each choice. BNN produces different answers each time, and the accumulation process can favor one choice over another until enough evidence is gathered for a decision.

The speed of decision-making was also tested, following a phenomenon known as the "speed-accuracy trade-off," which dictates that people make less accurate decisions when under time pressure. The results showed that the RTNet model mimics this phenomenon.

The researchers also found that RTNet behaves like humans in terms of decision confidence - people feel more confident when their decisions are correct, and RTNet showed similar characteristics without special training for it.

Future Research
The team plans to expand their research by training the network on more diverse datasets to test its potential. It is expected that this model will be applied to other neural networks to enable rationalization of decisions similar to humans. In the long term, algorithms could help alleviate the cognitive load of the thousands of decisions we make every day.

In addition, research at MIT is developing flexible networks, known as "liquid" neural networks, which adapt to changing conditions and enable better understanding and diagnosis of network decisions. These networks have shown high accuracy in predicting future values in various datasets, including atmospheric chemistry and traffic patterns.

Researchers at Stanford are exploring how neural networks can help with complex tasks such as predicting outcomes based on reward history and risk assessment, which could improve decision-making in unknown situations.

Ultimately, the goal is to develop algorithms that not only mimic our decision-making abilities but could even help reduce the cognitive load we carry every day.

Source: Georgia Institute of Technology