Figure 1.
Neural network implementation of the DDM and associated LRP predictions.
The neural network model of decision making consists of three layers (units with S-shaped curves denote bistable switch units). In the evidence accumulation layer (A), located potentially in parietal cortex, activity increases with motion input over time (red = leftward motion; blue = rightward motion). The difference in left minus right activations approximates a drift-diffusion process. This accumulation layer feeds into the threshold layer (B), which is potentially located in primary motor cortex. The difference between left and right switch unit activations then reflects the LRP signal. The threshold layer's output is transformed into a punctate motor output in the response layer (A), which also ensures that the threshold and accumulation layers are reset after the response. (D) Our model predicts that for changes in signal-to-noise ratio, in the accumulation layer the slope of neural activity changes, while in the threshold layer there is a non-linear change in slope. Consequently, the change due to increases in signal-to-noise ratio is best quantified by the area between curves. (E) Additionally, the model predicts that the magnitude and location of the response-locked simulated LRP should vary with signal-to-noise ratio (pink = low drift;green = high drift). (F) When varying response bias, the model predicts that the height of the LRP should change (pink = low bias;green = high bias).
Figure 2.
Behavioral performance in Experiment 1.
Mean (sem) coherence (A), accuracy (B), and response time (C) across subjects for the low and high coherence conditions. Coherences were tuned to ensure approximately 70 and 90% correct performance for each participant, and the two conditions are statistically different from each other for all three measures (coherence, accuracy, and response time).
Table 1.
Pure DDM parameters for best fitting model to data from Experiment 1.
Figure 3.
Response-locked LRPs and individual differences for Experiment 1.
(A) Grand average response-locked LRP, demonstrating the difference between low and high coherence conditions. Vertical lines indicate stimulus onsets for the respective conditions. Shaded area indicates the time window where low and high-coherence differ significantly (t-test with p<0.05). Inset shows a topographical map (nose up) of lateralized EEG activity, demonstrating that electrodes C3 and C4 are maxima of this measure. (B) Individual differences in DDM estimate of drift rate correlate with area between curves of high- and low-coherence LRPs. Each dot reflects the difference between low- and high-coherence drift and area between LRP curves for a single participant. (C) Grand average response-locked LRP demonstrating the difference between integration and non-integration conditions. Blue trace reflects the evidence-integration condition (average of low- and high-coherence trials). Red reflects a task condition where the participant has to press a pre-specified button, whereas green shows trials on which a participant is instructed by an arrow cue which button to press. Vertical lines indicate dot-motion onsets for the respective conditions. (D) We estimated non-decision time from the LRP by adding the time until departure from baseline to the distance between LRP peak and the actual motor response. The thus-estimated neural
correlates with the behaviorally-estimated
. Each dot reflects data from one participant in one condition (low or high coherence).
Figure 4.
Behavioral data for Experiment 2.
Mean accuracy (A) and reaction time (B) as a function of response bias condition. (C). Fitted DDM starting point increases with response bias condition. Response bias condition is operationalized as the proportion of trials in biased direction. Error bars reflect sem.
Table 2.
DDM parameters for best fitting model to data from Experiment 2.
Figure 5.
Response-locked LRPs in Experiment 2.
(A) Grand average LRP waveform separated by response bias (cf. Figure 1F). (B) Relationship between normalized LRP peak height and fitted DDM response bias. There is a significant correlation between the DDM starting-point parameter and LRP height. Error bars reflect sem. (C) Grand average response-locked LRPs in Experiment 2 separately for the preferred (solid) and non-preferred (dashed) direction. As predicted by our model, the LRP is larger for the non-preferred compared to the preferred direction [ = 11.4,
].