eLife—Wang Liping group discovered neural computing of causal inference in the frontal parietal loop of macaques

In daily life, we often receive a variety of sensory information
.
Properly integrating multisensory information can help us respond faster and more accurately [1].

Integrate or separate? This is the binding problem of information [2].

However, this process relies on our accurate inference
of the source of the information.
For example, when you cross the street, you will receive visual and sound information from the vehicle, both of which can reflect the
vehicle's location information to some extent.
When you infer that these two pieces of information come from the same source (the same vehicle), you integrate the two together for localization, otherwise you separate the information to locate
.
This process of inferring hidden causal structures based on observations is causal inference[3].

Causal inference is influenced by two factors: one is the influence of top-down prior information, such as your past experience, the street you are in; On the other hand, there are bottom-up effects, such as the spatial or temporal difference between the two sensory information [4].

Previous studies have shown that people integrate these two factors through Bayesian rules to obtain a posterior probability of common source (P_com) for a causal structure That is, the probabilistic inference of the causal structure [3, 5].

However, although previous studies have found that the computational process of causal inference obeys the bottom-up cortical hierarchy [2], there are still the following limitations: on the one hand, previous studies have mainly focused on human brain imaging, and lack the understanding of neural circuits at the level of single neurons; On the other hand, previous studies have focused on the neural relevance of behavioral representations, while how the brain represents hidden causal structures and how prior probability of common source (P_prior) and sensory representations are dynamically described remain unknown
.

On October 24, 2022, the Wang Liping Research Group of the Center for Excellence in Brain Science and Intelligent Technology (Institute of Neuroscience) of the Chinese Academy of Sciences published an online title in the eLife journal The research paper "Neural dynamics of causal inference in the macaque frontoparietal circuit" is the first to describe neural characterization and computational mechanisms
at the neural circuit level of causal inference.
PhD student Qi Guangyao and postdoctoral fellow Fang Wen are the co-first authors of the paper, and researcher Wang Liping and postdoctoral fellow Fang Wen are co-corresponding authors
.
This study uses a virtual reality system to perform electrophysiological recordings while awake macaques complete a causal inference task, and combines Bayesian modeling analysis and machine learning methods to first extend the original static causal inference model on the computational model, revealing how prior knowledge and sensory information in the causal inference model are updated with changes in the environment.
Secondly, it is found that (i) prefrontal motor cortex neurons characterize and update causal structures at the loop realization level to solve the binding problem of multimodal information; (ii) To maintain their consistency with the inferred environment, parietal cortical neurons receive premotor cortex feedback and update the characterization of sensory information to support the plasticity
of the body's sense of self-awareness.
The study discovered a new computational system for causal inference
.

In order to study the neural circuit characterization and computer system of causal inference, the researchers first designed a causal inference task based on a virtual reality system that can change the spatial position difference between visual and proprioceptive sensation (Figure 1A) [6].

In this task, macaques need to "report" the position of their arms based on the visual information and proprioception they receive (Figure 1B).

This task condition is called a Visual-proprioceptive conflict (VPC).
condition）
。 In this task, macaques infer a posterior probability (causal inference) that visual and proprioceptive information comes from the same source, and use this as a weight to estimate the position of the final arm [6].

In addition to VPC conditions, macaques require pure proprioception only (P) conditions and visual-proprioceptive congruence conditions （Visual-proprioceptive congruent (VP) condition）
。 Here, P and VP serve as reference baselines
for complete separation and full integration, respectively.

Figure 1: Results of behavioral tasks, dynamic causal inference models, and proprioceptive bias

(Source: Qi, et al.
, eLife, 2022).

In order to study the dynamic update of macaque behavior in the process of causal inference, combined with this experimental paradigm, the researchers proposed a dynamic causal inference model
.
This model adds an updated part of prior knowledge and proprioception to the previous Bayesian causal inference model (Figure 1C).

The model assumes the following:

First, macaques combine sensory information (visual and proprioception) and prior information in a Bayesian manner to judge arm position (Figure 1C).

Therefore, the behavioral output of macaques, that is, the distance of the arm subjectively reported by different visual differences (proprioceptive shift, Drift), will exhibit typical causal inference characteristics
.
For smaller disparity, P_com should be added (consolidated); When Disparity gets bigger, _Pcom should decrease (detach).

Second, macaques at the single trial level, the P of the current trial _{is subject} to the environment (the P of the previous trial _com and adjust the effect of current causal inference of prior information by updating P prior in the current state (Figure 1C1).

Third, in order to maintain consistency with a priori beliefs about causal structures, the early processing of sensory information is also influenced by the environment
.
When proprioceptive and visual information collide (e.
g.
, VPC tasks), the brain, in addition to reducing the prior belief that this information comes from a source (P_prior), also reduces the precision of sensory representations, thus balancing the credibility of current sensory signals and prior beliefs (figure).
1C3）
。

To test these hypotheses, the researchers used dynamic causal inference models to evaluate monkey behavior and studied the implementation
of neural circuits in multiple brain regions.

Consistent with the model's assumptions, the behavioral results of macaques show that: (i) the proprioceptive bias of macaques exhibits a nonlinear dependence on differences between proprioceptive and visual inputs, i.
e.
, when Disparity is smaller, macaques tend to integrate the two types of information, while when When Disparity is large, macaques tend to separate the two types of information (Figure 1D).

The causal inference model explains this well (Figure 1D, 2B), i.
e.
when Disparity is small, P_com Larger, macaques tend to integrate two types of information; While when Disparity is larger and P_com is smaller, macaques tend to separate the two types of information
.
(ii) P _priori is combined with sensory input and updated on a trial-by-trial basis based on historical information (P_com).
(Figure 2A).

(iii) To maintain consistency in causal inference, proprioceptive accuracy is updated during causal inference as _Pcom changes (Figure 2C).

。 The above results validate the model's hypothesis at the behavioral level that macaques adapt to environmental changes by integrating prior information and sensory inputs to infer hidden causal structures, while dynamically updating P prior and sensory representations.

This laid a theoretical foundation
for the subsequent study of the neural circuit mechanism of causal inference.

Figure 2: The causal inference model predicts the dynamic update of monkey behavior

(Source: Qi, et al.
, eLife, 2022).

Next, in order to study the neural mechanism of causal inference, the researchers recorded the premotor cortex (PMC) and parietal cortex (Parietal area 5), which are closely related to multisensory perception and body representation, during the macaque performing causal inference tasks and area 7) neuronal activity [7, 8].

Through information encoding and decoding analysis, the researchers analyzed the neural circuit mechanism
of dynamic computation of causal inference.
At the single-neuron level, the researchers found that task-related information was widely present in the frontopietal circuit
.
1) For sensory signals, parietal cortical neurons encode more proprioceptive arm position information, while PMC neurons encode more arm position information
after integration.
2) In addition to PMCs, a certain proportion (>5%) of the coding P_com is also present in the parietal cortex of neurons (causal inference neurons).

3) At the population neuron level, by combining machine learning algorithms to decode and analyze the task components and latent variable - P_com, the researchers found that the update of the hidden causal structure only appeared in PMCs , and its characteristic causal inference predates the parietal cortex, and then transmits signals to the parietal cortex through the communication subspace (Figures 3A and B); 4) The parietal cortex receives _Pcom information from PMCs and adapts to the environment by updating the encoding of proprioception (Figures 3B and C).

These results validate the hypotheses and behavioral observations of dynamic causal inference models from the perspective of neural activity
.

Figure 3: Neural implementation of dynamic causal inference

(Source: Qi, et al.
, eLife, 2022).