Cell's "dimensionality reduction blow" from neurons: the decision-making thinking framework of the parietal lobe

Written | Edited by Su Yiting | Xi We make decisions every day, from what to eat for lunch, to what major to enroll in university and so on
.

These decisions first require the sensory organs to collect external information, then analyze and integrate the external information, and finally make a decision.
The brain integrates the collected information into so-called decision variables (DV), by comparing the decision variables and whether they reach a certain value.
Threshold to make a decision
.

For example, for today’s lunch, if a restaurant’s taste is more than 7 points, you decide to patronize.
Restaurant A’s dishes have been eaten several times in the past.
At first, you stepped on the thunder for 6 points.
Later, the restaurant introduced new dishes that suit your taste.
Gradually, In fact, your evaluation of restaurant A is getting better and better, and you can probably get a score of 7.
5.
Since 7.
5>7, restaurant A is undoubtedly the object of your patronage today
.

Past studies believe that the firing frequency of neurons in certain areas of the brain can reflect decision-making variables.
For example, the firing frequency of neurons in the LIP brain area reflects the important decision-making information accumulated in the brain
.

In the sensory information collection stage before decision-making, preference for sensory information of a certain decision will cause the neurons encoding this decision module to become more active
.

In the final decision-making action stage, neurons of different decision-making modules compete with each other until one of the modules wins
.

What has not yet been answered in the scientific community is how do neural activities affect decision-making in the intermediate information integration stage? Is the information integration process the same between different tasks? On June 24, 2021, the Roozbeh Kiani team from New York University published an article Representational geometry of perceptual decisions in the monkey parietal cortex on Cell, and found out the representational structure of neurons in the decision-making process.
They believe that different tasks The integration of information between is different
.

In this study, researchers trained rhesus monkeys to complete two face categorization tasks (face categorization tasks, hereinafter referred to as facial movements) and one motion direction discrimination task (motion direction discrimination task).
At the same time, single-cell recording was used to capture the discharge activity of the neuron population in the lateral intraparietal (LIP) brain area during the task
.

It was found that the firing frequency of neurons in the facial task was negatively related to the difficulty of the task, while the response of the neurons in the motor task was exactly the opposite
.

If the group firing responses of neurons are projected into a three-dimensional principal component space, their manifold in the three-dimensional space will rotate with the change of the task.
Even in the same task, different rules will cause them to move in the three-dimensional space.
Presents differently
.

In the task of distinguishing the direction of movement, the experimental monkey needs to find the direction of the consistent movement of the random point in the dynamic random point visual stimulus, and get rewards by reporting the direction of the change in eye movement
.

The motion coherence of random points is distributed from strong to weak.
For example, a 50% consistency means that 50% of random points move in the same direction at the same time
.

In the face part task, the experimental monkey needs to distinguish whether the facial stimulus is a human face or a monkey face, or a happy face or a sad face
.

In order to create different task difficulty levels, the experimenters morphed the face to create a “face makeup” that gradually changed from a human face to a monkey face, a happy face to a sad face
.

In the decision-making stage, the experimental monkey finally made a choice between the left and right targets through eye movements, one of which was placed in the response field of the LIP nerve cell (Tin), and the other was placed in the response field (Tout)
.

The researchers found that the experimental monkeys had similar decision-making behaviors in the two tasks: the longer the visual stimulus stayed, the higher the accuracy of the task, and the accuracy also decreased with the increase in the difficulty of the task, indicating that the experimental monkey had one before the decision The process of information gathering
.

This behavior is also consistent with the simulation results of the Drift Diffusion Model (DDM)
.

Since the behavior of the two tasks is similar, are the responses of LIP neurons also similar? The answer is no
.

The researchers found that in the motor task, the firing frequency of LIP neurons in the case of Tin increases with the increase of the visual stimulus intensity (that is, the consistency of random points), and the firing frequency in the case of Tout is just the opposite.
, That is, the smaller the stimulation intensity, the higher the discharge frequency
.

In the face task, although the LIP neuron reacts the same as the motor task in the case of Tout, in the case of Tin, the lower the visual intensity of the facial stimulus, the higher the firing frequency-as if the more difficult it is to distinguish When the face is a human or a monkey, happy or sad, the more the brain tries to find the answer
.

This result tells us that in different tasks, the neural calculations in the background of the brain are likely to be different
.

In order to further understand the neural calculation principle behind this, the researchers performed principal component analysis on the discharge activity data of neuron populations, and finally projected them from the multidimensional space into the three most important dimensions-just like photography can transform the three dimensions in real life.
The thing is like a two-dimensional movie.
Although it is reduced by one dimension, the movie still retains most of the information
.

The presentation of data points of these neurons in this three-dimensional space (hereinafter referred to as state space) reflects the collective response of neurons to stimuli of different intensities at different times after the start of the visual stimulus
.

In this state space, the data points of the group neural response slowly change along a curve with different visual stimulus intensities (referred to as a curved manifold in the text)
.

This curved manifold continues to extend over time after the visual stimulus starts, slowly separating the presentation of different stimuli and different stimulus intensities
.

If you project this curved manifold on the discharge frequency axis, you can once again see the discharge frequency response of the facial task and the motor task to the stimulation intensity (that is, the greater the difficulty of the former, the more frequent the discharge vs the greater the difficulty of the latter.
The rarer)
.

The curved manifold of this space indicates that LIP neurons are likely to encode two kinds of information at the same time during the decision-making process: the difficulty of visual stimulation and the decision variable (DV) that affects decision-making behavior
.

The researchers found a pair of two axes that best represent the difficulty of stimulation and decision variables in this space through analysis, so that these two axes can still well reflect the three-dimensional curve of neuronal population response in the two-dimensional space.
Degree manifold
.

Sure enough, they found that the stimulus difficulty axis and the decision variable axis can respectively predict the actual stimulus difficulty and the correctness of the decision
.

However, these two axes cannot predict the experimental monkey’s confidence in task judgment, nor can it predict the experimental monkey’s actual action plan.
It is likely that the latter two use different coding rules
.

Finally, in the two face recognition tasks, if the same visual stimuli are presented, only the rules are changed: one rule is to judge a human face or a monkey face, and the other is to judge happy or sad.
In this case, it is neurological.
Is the background calculation the same? The researchers did the same analysis on LIP neurons under these two rules, and found that as long as the task rules are different, their curvature manifolds in the state space are different, and their projection angles on the firing frequency axis are also different.
The same
.

This result shows that when the task sensory stimulus is exactly the same, the change of the game rules will also change the coding calculation of the neuron
.

One of the signs of decision-making is its flexibility.
In order to make flexible decisions, the brain must be able to grasp information in different contexts and understand the underlying rules to guide actions
.

This article analyzes the spatial curvature manifolds of LIP neurons in different tasks, which may provide new ideas for the occurrence of decision-making behaviors
.

Original link: https://doi.
org/10.
1016/j.
cell.
2021.
05.
022 Platemaker: Notes for reprinting on the 11th [Original article] BioArt original article, personal forwarding and sharing are welcome, reprinting without permission is prohibited, all published The copyright of the work is owned by BioArt
.

BioArt reserves all statutory rights and offenders must be investigated
.