关键词: 认知过程, 证据积累模型, 计算建模, 人工神经网络

Abstract:

Reaction time (RT) is a window into understanding human decision-making processes. The Evidence Accumulation Model (EAM) is a dominant computational framework for modeling RT. However, EAMs, such as the Drift Diffusion Model (DDM), offer statistical descriptions of decision outcomes without detailed algorithms for stimulus encoding or neural mechanisms, thereby omitting the algorithmic and hardware levels in David Marr’s three-level framework (computation, algorithm, and hardware). We suggest that these limitations can be addressed by combining Artificial Neural Networks (ANNs) and evidence accumulation models to simulate the entire decision-making process—from stimulus encoding to decision output. These new models, termed Cognitive Decision Neural Networks, enable in silico modeling of human decision-making, providing a novel approach to understanding cognitive processes.

Cognitive Decision Neural Networks generally consist of three key modules: stimulus encoding, evidence accumulation & decision-making, and reaction time output. The stimulus encoding module encodes sensory data into decision evidence, capturing task-relevant information. The evidence accumulation & decision-making module integrates task-relevant evidence, sets decision rules to make a choice and generates corresponding decision time, capturing both encoding and decision processes. The output module aligns model’s decision time with human reaction time, ultimately producing the final behavioral output.

Recently, a few models based on ANNs have combined with the evidence accumulation model to simulate speedy decision-making processes in several tasks. The RTNet employs Bayesian neural networks to model uncertainty and reaction time dynamics for handwriting recognition task. RTify, which integrates a recurrent neural network to simulate evidence accumulation under temporally evolving stimulus, fits the human perceptual decision-making RT data well. Both RTNet and RTify use deep neural networks to handle complex visual stimuli; however, spiking neural networks can also be used. For example, SN-DM (spiking neural decision-making model) implements biologically plausible neural coding through spiking neuron populations to investigate neural mechanisms underlying decision processes. All these examples suggest that ANNs provide a new tool for modeling the stimulus encoding, decision-making process, and the neural dynamics underlying the decision-making process.

While Cognitive Decision Neural Networks show promise in lightweight, interpretable modeling, their generalizability remains limited, particularly in complex tasks such as social or value-based decisions. Future research may explore advanced ANNs’ architectures (e.g., tiny RNNs) or hybrid evidence accumulation mechanisms to enhance flexibility. Integrating multimodal data (e.g., neural, eye-tracking) and embodied AI (e.g., robotic arms) may also improve the performance and help address issues such as the non-decision time. In sum, the Cognitive Decision Neural Network framework outlines new frontiers for modeling human decision-making and offers new insight for understanding human cognition.

Key words: cognitive process, evidence accumulation models, computational modeling, artificial neural networks