Rat Prefrontal Cortex Inactivations during Decision Making Are Explained by Bistable Attractor Dynamics

Abstract

Two-node attractor networks are flexible models for neural activity during decision making. Depending on the network configuration, these networks can model distinct aspects of decisions including evidence integration, evidence categorization, and decision memory. Here, we use attractor networks to model recent causal perturbations of the frontal orienting fields (FOF) in rat cortex during a perceptual decision-making task (Erlich, Brunton, Duan, Hanks, & Brody, 2015 ). We focus on a striking feature of the perturbation results. Pharmacological silencing of the FOF resulted in a stimulus-independent bias. We fit several models to test whether integration, categorization, or decision memory could account for this bias and found that only the memory configuration successfully accounts for it. This memory model naturally accounts for optogenetic perturbations of FOF in the same task and correctly predicts a memory-duration-dependent deficit caused by silencing FOF in a different task. Our results provide mechanistic support for a "postcategorization" memory role of the FOF in upcoming choices.

Figure 1:

Cortical inactivation produces an ipsilateral bias during decision making. Description of the behavioral data modeled in this study. (Left) Task schematic of an evidence accumulation task (adapted from Brunton et al., 2013; Erlich et al., 2015; Hanks et al., 2015). Rats enter a center nose port and hear Poisson-generated clicks from both a speaker to their left and a speaker to their right. After the click trains have ended, the rats must enter a nose port on the side that played the greater total number of clicks to get a reward. (Right) Muscimol infusion into the frontal orienting fields produces an ipsilateral bias. Sensory instructed trials in which a visual sensory signal (LED turning on in the reward port) indicates which of the two side ports is the correct choice, are not biased after FOF inactivation (left/right LED). (Reproduced from Erlich et al., 2015.)

Figure 2:

Conceptual stages of decision making. (A) We consider three stages of decision making that might be perturbed during FOF inactivation: integration, decision categorization, and decision memory. In this cartoon, seven clicks are presented on the left and four on the right. The evidence is integrated, categorized into a go-left trial, and the decision is remembered. (B) Schematic of the integration model. Two nodes represent populations of neurons that self-excite and have cross-inhibition. Each node gets feedforward input consisting of the evidence click trains. (C) Schematic of the categorization and postcategorization models. The accumulated evidence a is passed through a thresholding function f () that has a parameter g describing how “soft” or “hard” the thresholding is and then used as inputs into the model. In the categorization model, the parameter g is fixed such that the function f () is linear, so that categorization of the value of a into a left or right choice occurs within or after the mutual inhibition model. In the postcategorization model, the parameter g = 0; then f () becomes the Heaviside step function, indicating the choice has already been categorized before entering the mutual inhibition model. When fitting the postcategorization model, we allowed g to be a free parameter in order to fit the degree to which inputs have already been categorized.

Figure 3:

The integration model fails to fit the data; The postcategorization model fits the data well. (A-C) Solid lines indicate model behavior. Dashed lines indicate rat data. Data from left and right inactivations from all rats have been transformed into one “meta-rat” data set. Ipsilateral refers to the side of muscimol infusion relative to choice. Error bars on data indicate 95% confidence intervals. (A) The best-fit integration model on a test set. (B) The categorization model (setting g = 50) fails to match the bias on easy contralateral trials. (C) The best-fit postcategorization model (fitting g) on a test set matches both the control and perturbed psychometric curves. (D) A visualization of P(a|Δ_R,_L). For every click difference, the average distribution of accumulator values predicted by the accumulation model is shown as a heat map. (E) Possible thresholding functions. Changing g allows the threshold function f () to vary from linear to the Heaviside function. Five example values of g are plotted here: 0,2, 5,10, and 50. The minimum and maximum accumulation values used to bound the domain of f () were determined by the accumulation model and fit to data. (F) The learned threshold function f () for the memory model is a steep sigmoid function (g < 1). This indicates the input into the model is already categorical, and thus the FOF supports a postcategorization memory.

Figure 4:

Dynamics of postcategorization model (A) Phase plane diagram for the postcategorization model during the memory period without inactivation. Solid lines are the nullclines of the external variables. (B) Same as panel A, but with partial muscimol inactivation of the left node. (C) Bifurcation diagram with respect to the inactivation fraction h. Near h = 0.35, the left attractor disappears in a saddle node bifurcation. Lines marked A and B show the parameter values used in panels A and B. (D) Mean activation of each node during a Go-Left trial (a = −10).

Figure 5:

Properties of postcategorization model. (A) Tuning curve for FOF model showing the firing rate as a function of the accumulation value. Constructed using the method developed in Hanks et al. (2015). The slope of the tuning curve at a = 0 is 0.133, comparable to the FOF population average of 0.158 ± 0.015 reported in Hanks et al. (2015). (B) Psychometric curve for bilateral muscimol inactivations of the FOF. The model (solid line) is able to match the psychometric curves from the data (dashed). The model was fit simultaneously to both unilateral and bilateral data sets. (C, D) The postcategorization model was simulated using temporally precise inactivations, mimicking opto-genetic experiments. Error bars show model variability over repeated sampling of trajectories. Three trial epochs were inactivated: the first and second half of the evidence period (yellow, green) and the memory period (blue). All trials had a 1 s evidence period. Memory period duration was 100 msec. Inactivation effect is shown as the percentage of bias on model full trial inactivation. (C) Unilateral inactivation produces an ipsilateral bias. (D) Bilateral inactivation, showing average impairment on each side.

Figure 6:

Postcategorization bias grows with memory period duration. (Left) Model prediction for bias during unilateral muscimol inactivation of the FOF as a function of increasing memory period duration. The model produces a bias that increases over time. Bias is relative to control behavior at each memory duration. The model was fit to the Poisson clicks task, not the MGO task. (Right) Unilateral FOF inactivation data from Erlich et al. (2011) shows a bias that grows with memory period duration for MGO trials. Trials were either nonmemory or had a memory period sampled from a distribution with mean 750 msec. Memory period is the time from the end of the stimulus to the go signal. Both trial types had an additional effective memory period between the go signal and the rat’s response. Memory trials were binned by their memory durations into shorter or longer than 750 msec. The solid line shows the average bias on ipsilateral trials at each of the three time bins. Bias in each time duration is relative to control performance on the same time duration. Nonmemory trials show some bias, while longer memory periods have stronger biases. Note that the vertical axis scale is different between plots.