Recently, researchers from Georgia Institute of Technology developed RTNet, demonstrating for the first time that its "way of thinking" is very similar to humans.
In terms of capabilities, current AI's expertise has surpassed humans in many aspects. However, humans still retain some unique characteristics, such as the brain's efficiency, emotions, and soul. Whether superintelligence needs to learn these human characteristics may require further exploration.
The RTNet developed by researchers is the first neural network with a thinking process similar to humans. Traditional neural networks' decision-making behavior differs significantly from humans, while RTNet can simulate human perceptual behavior, generate random decisions, and produce response time distributions similar to humans.
RTNet's internal mechanism is closer to the actual mechanism of human response time generation. Its core assumption is that response time is generated by a process of sequential sampling and result accumulation. The network structure is divided into two stages:
-
The first stage uses the Alexnet architecture, but with weights in BNN form, randomly sampling weights from learned distributions during each inference, introducing randomness.
-
The second stage is an accumulation process, accumulating inference results until a threshold is reached.
RTNet simulates two characteristics of human decision-making in principle: the randomness introduced by BNN, and different completion times for tasks of varying difficulty. The authors demonstrated through comprehensive testing that RTNet replicates all fundamental features of human accuracy, response time, and confidence.
Human perceptual decision-making has six basic features:
- Decisions are random
- Speed pressure shortens response time but reduces accuracy
- More difficult decisions lead to lower accuracy and longer response times
- Response time distributions are right-skewed and increase with task difficulty
- Response times for correct trials are lower than for incorrect trials
- Confidence is higher for correct trials than for incorrect trials
The experimental design included a human control group and multiple neural network models including RTNet. The human control group performed a digit discrimination task, reporting perceived digits and decision confidence. The experiment tested speed-accuracy trade-off (SAT) and different task difficulties.
RTNet uses the Alexnet architecture with Bayesian Neural Networks (BNN) to introduce randomness. Researchers trained RTNet for 15 epochs, achieving over 97% classification accuracy on the MNIST test set.
The experiment also included other neural network models such as CNet and BLNet for comparison. CNet is based on a residual network architecture, using skip connections to introduce propagation delays. BLNet is an RCNN composed of a standard feedforward CNN and recurrent connections.