Summary: Researchers at Georgia Tech have developed a neural network that models human decision-making by integrating uncertainty and evidence accumulation. Trained on handwritten digits, this model—called RTNet—produces stochastic, human-like decisions, matching human patterns of accuracy, response time, and confidence more closely than traditional deterministic networks.
RTNet’s approach blends a Bayesian neural network with an evidence accumulation mechanism so the model gathers and weighs information over time before committing to a choice. When tested on noisy handwritten digits, RTNet not only matched human performance on average accuracy but also reproduced the characteristic speed-accuracy trade-offs and confidence patterns seen in people. This advance points toward AI that behaves more reliably and transparently in perceptual decision-making tasks.
Key Facts:
- Human-like decisions: RTNet mimics human uncertainty and evidence accumulation, producing variable responses instead of deterministic outputs.
- Performance comparison: When evaluated on noisy MNIST images, RTNet generated accuracy, response time, and confidence distributions similar to those measured in human observers.
- Future potential: Incorporating human decision signatures into neural networks may improve AI reliability and help reduce cognitive load by supporting or automating routine perceptual choices.
Source: Georgia Institute of Technology
Every day people make thousands of small decisions—studies estimate roughly 35,000—ranging from crossing a street to selecting a meal. These choices rely on accumulating sensory evidence, recalling past experience, and estimating confidence. Importantly, the same person may choose differently in identical situations at different times because of uncertainty and changing evidence.
Conventional neural networks are deterministic: given the same input, they repeatedly produce the same output. A team in Associate Professor Dobromir Rahnev’s lab at Georgia Tech has taken a different route, designing a model that embraces the variability of human decision-making to produce more natural, interpretable behavior.
Applying principles from human perceptual decision-making to machine learning is a relatively new direction, but it has the potential to yield models that behave more like people. The team reports their work in the journal Nature Human Behaviour in a paper titled “The Neural Network RTNet Exhibits the Signatures of Human Perceptual Decision-Making.”
Why confidence and uncertainty matter
“Traditional neural networks typically produce a prediction without indicating how confident they are,” explained Farshad Rafiei, who completed his Ph.D. in psychology at Georgia Tech. That difference has practical consequences: large language models and other AI systems may generate plausible-sounding but incorrect answers without signaling uncertainty, whereas humans often acknowledge when they lack confidence or knowledge.
A model that reflects human uncertainty could reduce such errors and make AI behavior easier to interpret and trust.
How RTNet was built and tested
The researchers trained RTNet on the standard MNIST dataset of handwritten digits. To evaluate robustness and human similarity, they tested the trained model on a noisy version of the dataset that made digits harder to recognize. For comparison, they also evaluated three alternative architectures—CNet, BLNet, and MSDNet—training all networks on clean MNIST and testing on the noisy images used in human experiments.
RTNet combines two key mechanisms: a Bayesian neural network (BNN) that generates probabilistic outputs, and an evidence accumulation process that integrates support for each choice over time. Because the BNN outputs vary from pass to pass, RTNet’s decisions are stochastic. The accumulation process stops and makes a choice once the evidence for an option exceeds a threshold, mirroring how humans decide under uncertainty.
The team also measured decision time to test whether RTNet exhibits the well-known speed-accuracy trade-off: humans tend to sacrifice accuracy when forced to decide more quickly. Sixty Georgia Tech students participated in the behavioral experiment, classifying the same noisy images and reporting their confidence on each decision. RTNet’s distributions of accuracy, response time, and confidence closely matched the human data.
Importantly, RTNet outperformed the deterministic alternatives across multiple measures and reproduced human-like patterns—for instance, showing higher confidence on correct trials without being explicitly trained to do so. The model was also better at maintaining accuracy under faster decision constraints, reflecting a core feature of human psychology.
Outlook
The researchers plan to train RTNet on more diverse datasets and explore how BNN-based approaches can be incorporated into other architectures to promote more human-like reasoning. In the longer term, such models may not only emulate human perceptual decision-making but also help alleviate the cognitive burden of routine choices by providing calibrated, confidence-aware recommendations.
About this artificial intelligence research news
Author: Tess Malone
Source: Georgia Institute of Technology
Contact: Tess Malone – Georgia Institute of Technology
Image: The image is credited to Neuroscience News
Original Research: Closed access. “The neural network RTNet exhibits the signatures of human perceptual decision-making” by Dobromir Rahnev et al., Nature Human Behaviour.
Abstract
The neural network RTNet exhibits the signatures of human perceptual decision-making
Convolutional neural networks have advanced models of biological vision, but conventional networks differ from human decision behavior because they are deterministic and allocate the same computation to easy and difficult inputs. These differences limit their ability to model human perceptual behavior. RTNet addresses these gaps by generating stochastic decisions and producing human-like response time distributions. Comprehensive tests show RTNet reproduces foundational features of human accuracy, response time, and confidence, outperforming current alternatives.
To validate RTNet’s predictions on novel images, the authors collected accuracy, response time, and confidence data from 60 human participants performing a digit discrimination task. RTNet’s outputs for individual images correlated with the same measures recorded from humans. Participants whose behavior was closer to the human average were also closer to RTNet’s predictions, indicating the model captures average human performance. Overall, RTNet offers a promising framework for modeling human response times and the critical signatures of perceptual decision-making.