Evidence accumulation occurs locally in the parietal cortex

Zhewei Zhang, Chaoqun Yin, Tianming YangNature Communications 2022 · DOI

TL;DR

This study investigates where in the brain evidence accumulation for decision-making originates. By recording from neurons in the posterior parietal cortex (LIP) of monkeys performing a complex probabilistic reasoning task, the researchers found that LIP not only accumulates evidence but also performs the crucial initial step of transforming sensory information into usable evidence. This suggests that evidence accumulation, a fundamental cognitive process, is computed locally within the parietal cortex.

What Problem Does This Paper Address?

Decision-making, especially under uncertainty, often involves accumulating evidence over time until a threshold is reached, triggering a choice. While neural correlates of this "evidence accumulation" have been observed in various brain regions, including the lateral intraparietal area (LIP), it remained unclear where this accumulation process truly originates. Specifically, it was unknown if LIP merely receives an already computed accumulation signal from elsewhere or if it actively computes this signal itself. This distinction is critical for understanding the functional architecture of decision-making circuits. The challenge lies in disentangling the accumulation process from the preceding computations that transform raw sensory input into the evidence that can be accumulated.

Methods

The researchers designed a novel two-stage probabilistic reasoning task. In this task, monkeys had to choose between two targets (red or green) based on a sequence of visual stimuli. Crucially, each stimulus had a color and a shape, and its shape was associated with a weight that influenced the reward probability of targets of the same color. This design forced the monkeys to first process the sensory information (shape and color) to determine the "accumulable evidence" before they could actually accumulate it.

They recorded the activity of single neurons in LIP while the monkeys performed this task. To analyze the neural data, they employed a Poisson generalized linear model (GLM). This statistical model allowed them to decompose the neural responses into different components, specifically identifying:

  1. Early component: Reflecting the transformation of sensory inputs into accumulable evidence. This involved analyzing how neurons responded to intermediate variables like shape weight and color consistency/evidence.
  2. Late component: Reflecting the accumulation of evidence and the formation of the final decision, characterized by the typical ramping activity observed in LIP.

By analyzing the timing and nature of these components, they could infer whether LIP was involved in the initial processing stages or just the final accumulation.

Datasets

The study used data collected from two adult male rhesus macaques (monkeys M and H) performing the custom-designed probabilistic reasoning task. No pre-existing datasets were used for training or evaluation.

Key Insights and Ideas

  • LIP is a "one-stop shop" for evidence accumulation: The most significant finding is that LIP neurons don't just accumulate pre-processed evidence; they also perform the crucial step of transforming raw sensory information (shape and color) into the evidence that can be accumulated.
  • Dissociable computational stages: The neural activity in LIP could be clearly separated into an early phase (sensory-to-evidence transformation) and a late phase (evidence accumulation and decision formation). This temporal dissociation strongly supports the idea of local computation.
  • Two routes to evidence computation: The study revealed that monkeys might use different strategies to compute evidence. Monkey M seemed to rely more on "color consistency" (whether the stimulus color matched the target color), while Monkey H leaned towards "color evidence" (combining shape weight and color directly). This highlights the flexibility of neural computation.
  • Negative encoding of shape weight: Surprisingly, LIP neurons often showed suppressed responses to shapes with positive weights (which favored a target). This negative encoding suggests that the information is processed through intermediate stages or originates from brain areas with specific encoding schemes, rather than being a direct reflection of reward value.

Key Results

  • Timing matters: The transformation of sensory input into accumulable evidence (early component) occurred significantly earlier in LIP neurons than the evidence accumulation itself (late component). For example, in Monkey M, shape weight encoding latency was around 212 ms, while evidence encoding latency was around 682 ms.
  • Evidence for local computation: The presence of both the sensory-to-evidence transformation and the evidence accumulation stages within the same LIP neurons strongly supports the hypothesis that evidence accumulation is computed locally in LIP.
  • Monkey-specific strategies: Monkey M showed significant representation of color consistency (47/115 neurons), while Monkey H showed robust population-level encoding of color evidence.
  • Coupled representations: In individual neurons, the encoding of shape weight, color consistency, and accumulated evidence were often correlated, suggesting that the same neuronal populations handle these related computations.

Limitations and Future Work

  • Monkey-specific routes: While the study identified different computational routes used by the two monkeys, the exact neural mechanisms underlying these differences remain to be fully elucidated.
  • Origin of sensory input: The paper speculates that the initial sensory information might originate from areas like the orbitofrontal cortex (OFC), but direct evidence for this pathway is not presented. Future work could investigate the upstream sources of these signals.
  • Negative encoding explanation: The reason for the negative encoding of shape weights is not fully explained and warrants further investigation into the connectivity and processing within the broader neural network.
  • Generalizability: While the findings strongly suggest local computation in LIP for this specific task, future research could explore whether this holds true for other types of decision-making tasks and other brain areas involved in evidence accumulation.

Summary updated

Abstract

Decision making often entails evidence accumulation, a process that is represented by neural activities in a network of multiple brain areas. Yet, it has not been identified where exactly the accumulation originates. We reason that a candidate brain area should both represent evidence accumulation and information that is used to compute evidence. Therefore, we designed a two-stage probabilistic reasoning task in which the evidence for accumulation had to be first determined from sensory signals orthogonal to decisions. With a linear encoding model, we decomposed the responses of posterior parietal neurons to each stimulus into an early and a late component that represented two dissociable stages of decision making. The former reflected the transformation from sensory inputs to accumulable evidence, and the latter reflected the accumulation of evidence and the formation of decisions. The presence of both computational stages indicates that evidence accumulation signal in the parietal cortex is computed locally.
BibTeX
@inproceedings{zhang2022evidence,
  author = {Zhewei Zhang and Chaoqun Yin and Tianming Yang},
  title = {{Evidence accumulation occurs locally in the parietal cortex}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-32210-6},
}
Figure