Dynamic task-belief is an integral part of decision-making

Abstract

Natural decisions involve two seemingly separable processes: inferring the relevant task (task-belief) and performing the believed-relevant task. The assumed separability has led to the traditional practice of studying task-switching and perceptual decision-making individually. Here, we used a novel paradigm to manipulate and measure macaque monkeys' task-belief and demonstrated inextricable neuronal links between flexible task-belief and perceptual decision-making. We showed that in animals, but not in artificial networks that performed as well or better than the animals, stronger task-belief is associated with better perception. Correspondingly, recordings from neuronal populations in cortical areas 7a and V1 revealed that stronger task-belief is associated with better discriminability of the believed-relevant, but not the believed-irrelevant, feature. Perception also impacts belief updating; noise fluctuations in V1 help explain how task-belief is updated. Our results demonstrate that complex tasks and multi-area recordings can reveal fundamentally new principles of how biology affects behavior in health and disease.

Keywords: area 7a; cognitive flexibility; decision-making; electrophysiology; macaque monkey; parietal cortex; primary visual cortex; recurrent neural network; rule switching; visual cortex.

Figure 1.. Behavioral paradigm and electrophysiological recording.

(A) Schematic of the two-interval, two-feature discrimination task with stochastic task switching. On each trial, monkeys discriminate the difference in either spatial frequency or spatial location between two subsequent Gabor stimuli. The relevant feature is uncued and changes with 2.5% probability on each trial. The monkeys indicate their perceptual decision and the feature believed to be relevant by making a saccade to one of four choice targets. (B) and are rewarded for correctly reporting the sign of the change in the relevant feature. (C) Belief-based decisions could potentially be solved by independent hierarchical processes that compute belief and perception (black boxes). While the animals do the task, we simultaneously recorded population activity from one representative brain region related to each process (7a and V1 respectively, blue squares show approximate implant locations) to test the hypothesis that these processes are not separable (red arrow).

Figure 2.. Trained behavior of animals and recurrent neural networks.

(A) Psychometric curves showing the monkeys’ perceptual choice proportion as a function of spatial frequency (left panel) and spatial location (right panel) differences. The flat curves for the irrelevant feature show that animals successfully ignored irrelevant visual information. (B) Distribution of number of trials it took the monkeys to adapt to task changes across experimental sessions (mean 3.1 trials). (C) Distribution of number of trials the monkey took to adapt to the task change relative to an ideal observer model (see STAR Methods) (mean 1.5 trials). Positive values refer to occasions where monkeys were slower than the model; negative values indicate that the monkeys were accidentally faster. (D) Schematic of the inputs and outputs of the network. The network receives sequential inputs containing information about stimulus changes, past model choices, and choice feedback (reward history) over the course of the last several trials. Like the monkey, the network model is trained to infer the implicit task rule from recent history and make corresponding decisions. (E) Similar to the monkeys’ behavior in (A), the choices of a trained network are informed by the believed relevant feature information (solid line), while independent of the believed irrelevant feature information (dashed line). (F) Distribution of differences between reward rates between RNN and the monkeys across experiment session. Overall the RNN obtained more reward than the monkeys in the same experiment sessions. (G) Comparing perceptual performance following rewarded trials (abscissa) and unrewarded trials (ordinate) for monkeys (upper panel) and the recurrent network model (lower panel). Each point represents one stimulus condition of an experimental session, and we compute perceptual performance based on the subjectively chosen task, regardless of whether that task-belief was correct. Upper panel, the monkeys’ perceptual performance is better following rewarded trials than unrewarded trials. The distribution lies significantly below the unity line (p<10⁻⁶ for both monkeys and both features), showing lower perceptual performances following a non-rewarded trial than following a rewarded trial, with the same perceptual difficulty. Lower panel, the perceptual performance of the artificial network does not significantly depend on feedback history (p>0.05 for both features). (H) The difference between the monkey and network model is not explained by different extent of training. The upper panel shows that the loss function decreased during time, indicating the gradual learning process of the model on the task. In the lower panel, the black line indicates the network model’s difference between perceptual performance following rewarded and unrewarded trials at each point of training. The red line indicates the corresponding difference in the monkeys’ perceptual performance, which always lies outside the 90% confidence interval for the model (gray shading).

Figure 3.. Neuronal measures of belief strength.

(A) In a high dimensional neuronal space expanded by the activity of 7a units during the delay period, we find the best hyperplane to discriminate the task the animal performed on the trial. We define our single-trial neuronal measure of belief strength as the Euclidean distance from 7a population activity on each trial to the hyperplane. (B) Task-belief decoded from a neuronal population. Task-belief can be better classified by a linear hyper plane from area 7a population activity following rewarded than unrewarded trials. The abscissa and ordinate of each point show the performance of a linear classifier following rewarded and unrewarded trials respectively, with leave-one-out cross-validation. Each point represents one experimental session. (C) Belief strength is schematized as the distance from a rolling ball to a boundary. for trials leading up to the animals’ decision to switch tasks, the average belief strength decreased monotonically, changed sign right at the point the monkey decided to switch tasks and recovered as the new task-belief was reinforced (histograms in bottom panel). Normalized activity of task-selective 7a units tracked the same dynamics as decoded belief around task switches (lines in bottom panel). Error bars indicate standard errors.

Figure 4.. Belief and perception are linked on a trial-by-trial basis.

(A) Using a procedure similar to that described in Figure 3A, we define the perceptual discriminability of each stimulus feature change on each trial as the Euclidean distance from V1 population activity to the hyperplane that best classifies the stimulus change of that feature (e.g., higher vs. lower spatial frequency). (B) Trial-by-trial comparison between belief strength (abscissae, decoded from 7a) and perceptual discriminability (ordinates, decoded from V1) for an example stimulus/task condition. If belief decisions and perceptual decisions are implemented by separate functional circuits of the brain, then internal fluctuations of the two systems should have no correlation. (C) The belief- spatial location discriminability correlation is positive when spatial location is believed to be relevant (histogram and magenta cumulative distribution curve, p=4×10⁻⁶), but not when it is believed to be irrelevant (cyan cumulative distribution curve, p>0.05). The two distributions are significantly different (Wilcoxon rank sum test, p=0.014). (D) Similarly, belief- spatial frequency discriminability is significantly positive when spatial frequency is believed to be relevant (p=0.0015) but not when it is believed to be irrelevant (p>0.05). The two distributions are significantly different (Wilcoxon rank sum test, p=0.03).

Figure 5.. Trial to trial variability in visual cortex affects belief updating.

(A) Example neurometric curve showing the ability of a decoder to discriminate spatial frequency changes from the population of recorded V1 neurons using logistic regression on the perceptual discriminability of spatial frequency (as in Figure 4A). (B) Based on perceptual confidence on each trial (estimated from V1 population activity), a normative model determines whether the subject should switch tasks given the trial history (see STAR Methods). (C) Based on the V1 projections to the relevant feature subspace on each trial, we estimate from the neurometric curve, which represents the probability the monkey’s behavioral choice is correct (i.e. perceptual confidence (Hangya, Sanders and Kepecs, 2016)). In the trial-shuffle analysis, we randomly switch the confidence within trials with the same conditions (dots with same color). (D) Model predictions after trial-shuffle, conventions as in (b). (E) Trial-to-trial variability in V1 is related to belief. The models’ sensitivity indices (d’) measures the performances predicting whether the monkey would switch tasks in a session. The histogram shows the differences in the d’s of models using the actual and trial-shuffled V1 activity. Across experimental sessions, shuffling V1 activity among trials with identical conditions limited the model’s capability to predict the monkeys task-switching behavior (p=5×10⁻⁴).