Evidence weighting in uncertain and correlated environments.

Our understanding of how the brain makes decisions has benefited greatly from behavioral and neural findings that have identified close links between neural computations and normative decision theory. Among the most well-studied decisions that exhibit this link are simple perceptual decisions that are thought to transform sensory observations into likelihoods, which are the probabilities of obtaining the sensory evidence given particular hypotheses about the state of the world (Gold & Shadlen, 2001, 2007). According to normative theory, likelihoods for different hypotheses can be compared to each other (e.g., via their ratio) to make optimal decisions (in terms of maximizing accuracy) about which hypothesis about the world is most probable given the evidence. This procedure is central to numerous models, such as the widely used drift-diffusion model (DDM), that have successfully accounted for a host of decision-related neural and behavioral findings.

However, many of these models include “fudge factors” that are used to account for the fact that true likelihoods can be difficult to compute and thus must often be approximated. This difficulty arises because true likelihoods require information about not just the given sensory observation but also the (possibly long-term) statistics of the process that generated that observation in the world and then represented it in the brain. Previous studies have shown that these fudge factors (e.g., the drift rate in the DDM, often fit as a free parameter) can vary considerably across individuals and conditions. Because these fudge factors govern the difference between true likelihoods, which are needed for optimal decisions, and merely approximate ones, this variability can play a key role in governing decision accuracy, speed, biases, and other factors related to optimality. Although these sub-optimalities have been identified for certain conditions, a more general treatment, including identifying how different statistical environments affect decision-making via their impact on how (and how easily) likelihoods can be computed or approximated from sensory observations, is lacking and is the focus of this study.

In particular, most prior work focused on decision-making based on single streams of evidence or based on multiple sources that were assumed to be statistically independent. However, in the real world, we often must base decisions on evidence that comes from multiple sources that are not independent. Ignoring correlations among sources of evidence distorts the computation of likelihoods and posteriors, leading to them to over- or underestimate the probability of particular hypotheses. When accumulating evidence over time to form decisions based on a fixed decision rule, such distortions result in unintended shifts of the speed-accuracy tradeoff. However, the extent to which human decision-makers take correlations into account is not well-understood.

Here we developed a novel evidence accumulation paradigm that allows precise control of the correlations between sources of evidence to ask whether and how humans take correlations into account in simple two-alternative perceptual decisions.

我们对大脑如何做出决策的理解，得益于行为学和神经科学的研究成果，这些成果揭示了神经计算与规范性决策理论之间的紧密联系。其中，最被广泛研究且显示出这种联系的决策类型是简单的感知决策，这类决策被认为是将感官观察转化为可能性，即在世界状态特定假设下获得感官证据的概率（Gold & Shadlen, 2001, 2007）。根据规范性理论，不同假设的可能性可以相互比较（例如，通过它们的比率）以做出最优决策（从最大化准确性的角度），即根据证据判断哪个关于世界的假设最有可能。这一过程是众多模型的核心，如广泛应用的漂移扩散模型（DDM），这些模型成功地解释了众多与决策相关的神经和行为学发现。

然而，许多这些模型包括了所谓的“调整因子”，这些因子被用来解释真实可能性难以计算的事实，因此往往需要近似。这种困难源于真实可能性不仅需要关于给定感官观察的信息，还需要生成该观察在世界中的过程及其在大脑中的表征的（可能长期的）统计信息。先前的研究表明，这些调整因子（例如，DDM中的漂移率，通常作为一个自由参数进行拟合）在个体和条件之间可能会有很大的差异。由于这些调整因子决定了真实可能性（这是做出最优决策所必需的）与仅是近似值之间的差异，因此这种变异性可能在决定决策的准确性、速度、偏差以及其他与最优性相关的因素中发挥关键作用。尽管这些次优性在某些条件下已被识别，但一个更通用的处理方法，包括识别不同的统计环境如何通过影响从感官观察中计算或近似可能性的方式和难易程度来影响决策，仍然缺乏，这也是本研究的重点。

特别是，先前的大多数工作集中在基于单一证据流或基于被认为是统计独立的多个来源的决策上。然而，在现实世界中，我们常常必须基于来自多个来源且这些来源并非独立的证据做出决策。忽视证据来源之间的相关性会扭曲可能性和后验概率的计算，导致对特定假设的概率高估或低估。当根据固定的决策规则积累证据以形成决策时，这些扭曲会导致速度-准确性权衡的不当转变。然而，人类决策者考虑到相关性的程度尚不清楚。

在此，我们开发了一种新颖的证据积累范式，该范式允许精确控制证据来源之间的相关性，以探讨人类在简单的双选择感知决策中是否以及如何考虑相关性。