Locus Coeruleus and Attentional Blink (Nieuwenhuis, et al, 2005)¶

“The Role of the Locus Coeruleus in Mediating the Attentional Blink: A Neurocomputational Theory”

Overview¶

This model implements a hypothesis concerning the role of the locus coeruleus (LC) in mediating the attentional blink. The attentional blink refers to the temporary impairment in perceiving the second of two targets presented in close temporal proximity.

In the attentional blink paradigm, each trial consists of a rapidly presented sequence of stimuli, one of which is marked as a target (e.g., by being a different color, or by being a digit rather than a letter). The participant is asked to observe the sequence, then identify the target (T1) as well as the stimulus that immediately follows it (T2). A consistent finding is that accuracy with which both T1 and T2 are identified depends on the lag between them. In particular, accuracy is high when T2 is presented immediately after T1, drops when T2 is presented between 200 and 300ms after the onset of T1, and is restored at delays longer than 400ms.

The model described by Nieuwenhuis et al. (2005) suggests that these findings can be explained by phasic activation of LC in response to T1, and the concomitant effects of norepinephrine (NE) release on neural gain, followed by refractoriness of the LC/NE system after a phasice response. The model demonstrates that accuracy in identifying T2 depends on:

whether T1 was accurately identified;

the lag between T1 and T2;

the mode of the LC

(phasic or tonic – see https://www.ncbi.nlm.nih.gov/pubmed/8027789).

The Figure below shows the behavior of the model for a single execution of a trial with a lag of 200ms (without noise), corresponding to the conditions reported in Figure 3 of Nieuwenhuis et al. (2005; averaged over 1000 executions with noise).

Nieuwenhuis et al. 2005 plot produced by PsyNeuLink

Nieuwenhuis et al. 2005 plot produced by MATLAB

The Model¶

The model is comprised of two subsystems: a behavioral network, in which stimulus information feeds forward from an input layer, via a decision layer, to a response layer; and an LC subystem that regulates the gain of the units in the decision and response layers. Each of the layers in the behavioral network is implemented as a pathway of TransferMechanism and LCAMechanism Mechanisms, and the LC subystem uses an LCControlMechanism and associated ObjectiveMechanism, as shown in the figure below:

Behavioral Network Subsystem¶

INPUT LAYER: a TransferMechanism with input_shapes=3 (one element for the input to the T1, T2 and distractor units of the DECISION LAYER, respectively), and assigned a Linear function with slope=1.0 and intercept=0.0.

DECISION LAYER: an LCAMechanism Mechanism of input_shapes=3 (one element each for the T1, T2 and distractor units), and assigned a Logistic Function with a slope=1.0 and intercept=0.0. Each element has a self-excitatory connection with a weight specified by self_excitation=2.5, a leak=-1.0, and every element is connected to every other element by mutually inhibitory connections with a weight specified by competition =1.0. An ordinary differential equation describes the change in state over time, implemented in the LCAMechanism mechanism by setting integrator_mode = True and time_step_size=0.02.

RESPONSE LAYER: an LCAMechanism Mechanism of input_shapes=2, with one element each for the response to T1 and T2, respectively, self_excitation=2.0, leak=-1.0, and no mutually inhibitory weights (competition=0).

PROJECTIONS: The weights of the behavioral network are implemented as MappingProjections. The matrix parameter for the one from the INPUT_LAYER to the DECISION_LAYER uses a numpy array with a value of 1.5 for the diagonal elements and a value of 0.33 for the off-diagonal elements; the one from the DECISION_LAYER to the RESPONSE LAYER uses a numpy array with 3.5 for the diagonal elements and 0 for the off-diagonal elements.

LC Subsystem¶

LC: an LCControlMechanism, that uses the FitzHughNagumoIntegrator to implement a FitzHugh-Nagumo model as a simulation of the population-level activity of the LC. The LCControlMechanism outputs three values on each execution:

v (excitation variable of the FitzHugh-Nagumo model) representing the state (i.e., net input) of the LC

w (relaxation variable of the FitzHugh-Nagumo model) representing noradrenergic output of the LC

\(gain(t)\), where \(g(t) = G + k w(t), G\) = base_level_gain, k = scaling_factor, and w(t) = the current noradrenergic output

The LC sends gain(t) to the DECISION LAYER and RESPONSE LAYER via ControlProjections in order to modulate the gain parameter of their Logistic Functions. Overall LC activity can be computed from v using the function \(h(v) = C * v + (1 - C) * d\), where C =0.90 and d = 0.5 (see Execution for additional details).

COMBINE VALUES: an ObjectiveMechanism, specified in the objective_mechanism argument of the LCControlMechanism constructor, with a Linear function of slope**=1 and **intercept**=0. Its **monitored_output_ports argument is assigned a 2-item tuple specifying the DECISION LAYER and a matrix for the MappingProjection from it to COMBINE VALUES. The matrix is assigned as a 3x1 numpy array, with weights of 0.3 for its first two elements and 0.0 for its third, corresponding to T1, T2 and distractor units in the DECISION LAYER, respectively. This combines the values of the T1 and T2 units, and ignores the value of the distractor unit, implementing the assumption that the distractor stimulus does not elicit an LC response. The COMBINED VALUES Mechanism conveys this combined value as the input to the LC.

Note

Although the COMBINED VALUES Mechanism is not strictly needed – the same MappingProjection and matrix used to combine the values of the DECISION LAYER and project to the COMBINE VALUES Mechanism could project directly to the LC (as it does in Niewenhuis et al.,2005) – the use of COMBINE VALUES conforms to the convention that PsyNeuLink ControlMechanisms are associated with an ObjectiveMechanism from which they receive their input.

Execution¶

The stimulus presentation sequence is split into 11 periods of execution, each of which is 100 time steps long. During each period, one of the three elements of the INPUT LAYER is assigned a value of 1 (activated) while the other two are assigned a value of 0. For the first three time periods, the distractor element is activated. T1 is activated during the fourth time period, followed in the fifther period by the distractor, and then by T2 during the sixth time period. During all other time period the distraction unit is activated. To reproduce Figure 3 of the Nieuwenhuis et al. (2005) paper, the log function is used to record the output values of parameters w and v for every execution of the LCControlMechanism. The function h(v) is computed for every time step, and h(v) and w are plotted.

Script: Download Nieuwenhuis2005Model.py