Humans engage in approximately 35,000 decision-making processes daily, ranging from mundane choices like deciding when it’s safe to cross the street to selecting what to eat for lunch. Each decision entails a complex process of evaluating options, recalling similar past experiences, and forming a decision with a certain degree of confidence. What is a quick, intuitive decision involves substantial environmental evidence collection. Moreover, the same individual may make different choices in identical situations at other times.

In contrast, traditional neural networks are deterministic, consistently yielding the same output under the same conditions. Under the guidance of Associate Professor Dobromir Rahnev, researchers at Georgia Tech are pioneering efforts to train neural networks to mimic human decision-making more closely. This approach of integrating human decision-making theories into machine learning is relatively new. Still, it promises to enhance the reliability of neural networks by aligning them more closely with human brain functions.

The team’s findings are detailed in a paper published in Nature Human Behaviour titled “The Neural Network RTNet Exhibits the Signatures of Human Perceptual Decision-Making.” This paper introduces a newly developed neural network trained to make decisions similar to humans.

One fundamental difference noted by Farshad Rafiei, who completed his Ph.D. in psychology at Georgia Tech, is that “Neural networks decide without indicating their confidence in the decision.” This contrasts sharply with human decision-making, where individuals typically acknowledge uncertainty. Addressing this discrepancy can prevent the inaccuracies and fabricated outputs—often seen in large language models (LLMs) when they encounter unknown queries—by fostering a design that admits uncertainty akin to human responses.

The neural network developed by the team was trained using the MNIST dataset of handwritten digits. To evaluate the accuracy of the network, the researchers tested it both on the original and a noise-added version of the dataset, simulating conditions that challenge human perception. They compared their model’s performance against human subjects and other neural networks, finding that their design adapted well to noisy data and mirrored human decision-making dynamics, including the psychological ‘speed-accuracy trade-off’—where quick decisions tend to be less accurate.

Key to their model’s functionality is a combination of Bayesian neural networks (BNNs), which apply probabilistic decision-making, and an evidence accumulation strategy that monitors the gathered evidence before making a decision. This setup allows the model to sometimes favour different outcomes based on the accrued evidence until a threshold is reached for a decision. The research also revealed that the RTNet mimics human confidence levels, often increasing certainty in correct decisions without specific training for this feature.

Additionally, Rafiei highlighted a significant gap in computer science literature regarding human behavioural data in response to visual stimuli, which limits the development of models that can accurately replicate human decision-making processes. Their work, contributing a substantial dataset of human reactions to the MNIST images, aims to mitigate this issue.

The researchers anticipate expanding their studies to include more diverse datasets and applying the BNN framework to other types of neural networks to enhance their human-like reasoning capabilities. The ultimate goal is to develop algorithms that simulate human decision-making and alleviate the cognitive load of individuals’ daily decisions. This pioneering work is a step toward creating neural networks that perform tasks and understand and interact in ways fundamentally akin to human beings.

More information: Farshad Rafiei et al, The neural network RTNet exhibits the signatures of human perceptual decision-making, Nature Human Behaviour. DOI: 10.1038/s41562-024-01914-8

Journal information: Nature Human Behaviour Provided by Georgia Institute of Technology