UC Irvine

The drift diffusion model (DDM) is a popular model of evidence accumulation that estimates parameters representing the underlying cognitive processes of decision making. DDM requires modeling the joint distribution of choice and response time (RT) with a Wiener first-passage time (WFPT) distribution to estimate a decision maker’s speed, caution, and bias, parameterized as drift rate, boundary separation and evidence starting point, respectively. Recent research demonstrates promising modeling results when parameters are allowed to vary across trials while being constrained by brain signals. We have developed Decision Sinc-Net, a novel, interpretable neurocognitive model that allows trial-level estimates of DDM by mapping EEG brain signals to the model parameters given behavioral data, through optimization of Wiener likelihood using gradient descent. Single-trial EEG data were used to represent the most likely cognitive parameters that gave rise to the observed choice response time. Critically, the lightweight neural network model was designed to automatically identify the neural correlates of different cognitive parameters in time and frequency domains without feature engineering. We showed that single-trial estimates of drift and boundary performed better at predicting RTs than median estimates did in both training and test datasets. This suggests that the model can successfully learn to extract meaningful trial-level EEG features to estimate Diffusion model parameters, and it generalizes well to out-of-sample brain data. We improved the model’s scientific interpretability by introducing an attention block to rank the learned frequency bands by importance. Interpretability methods were used to visualize how neural signals within the important frequencies were processed to estimate each parameter. The model was tested on two datasets. Results and architecture of the Decision SincNet model provided neural evidence for the DDM and demonstrated the possibility of an end-to-end neurocognitive modeling framework on a trial-level. To directly test the accumulation-to-threshold assumption in evidence accumulation, we introduced a probabilistic perceptual decision-making (pPDM) task to investigate computations of decision making processes through experimental manipulation. The task involves presenting stimuli as a succession of random samples with a known probability biased towards one of two alternatives. We quantitatively tested fundamental theories in sequential sampling during evidence accumulation. Hypothesis-driven models were used to evaluate which model best captures human behavior, under a Machine Learning framework for model comparison. Our results suggest that humans continuously integrate evidence, but multiple mechanisms are needed to explain the complex behavioral patterns observed. We developed two new models that best fit the data collected. One model is the integration of evidence weighted overtime. This model is similar to a discrete OU process with a recency effect, such that subjects focus on recent samples and suppress prior ones. The other model is a non-linear gradient boosted tree-based model that utilizes a combination of features: integration of evidence, number of samples seen, the order in which samples are presented, and pattern of recent samples. Our results suggest that computations during decision making consider all the factors as well as the interactions between them. Level of evidence and number of samples are the most dominant criterion. Subjects showed a higher probability of making a decision as evidence and the number of samples increased, but the relationships were both non-linear. Together, Decision SincNet and behavioral results from pPDM contribute to the theoretical understanding of human decision making through neurocognitive modeling and experimentation.