Adaptive population codes for flexible visually-guided behaviors

摘要/摘要 主动视觉需要根据当前任务目标对相关信息进行编码和记忆。 经典的观点认为，感觉编码、注意力选择和工作记忆是由 具有固定调谐曲线的视觉反应神经元的放电率或增益的持续变化 (称为“纯”或“固定选择性”神经元)。对固定选择性神经元中的增益调制的关注 揭示了大量关于注意力和记忆的基本机制。然而，它正变得越来越多 明确任务需求的动态变化可能需要更灵活的编码方案。例如，持有 工作记忆中与刺激诱发反应格式相同的信息可能会导致干扰 有新的感官输入。同样，灵活地编码感觉表征来完成一项任务--比方说一个简单的任务 在两种运动反应之间进行选择-可能需要重新配置表征，如果另一个刺激- 响应映射突然变得相关起来。最后，感官编码必须是灵活的，因为早期 在加工过程中，它们应该形成高维表示，以尽可能多地表示信息 关于当今世界的状况。在稍后的处理中，当需要做出决定或马达反应时， 代码应该折叠为只表示相关选项的较小子集。所有这些计算都是 更自然的实现是通过神经元的操作，神经元具有灵活的感觉功能和 对于任务需求(称为“混合选择性”神经元)。 基于这些考虑，我们假设灵活的行为是由混合选择性支持的 “旋转”高维神经编码的神经元对干扰变得健壮，或服务于其他 任务需求的变化。我们将使用建模、心理物理学和功能磁共振成像 (FMRI)来测试关于混合选择性应该如何调节大规模激活模式的预测 在人类受试者身上进行非侵入性测量。总的来说，这项工作将挑战传统的感官理论 基于固定选择性概念的编码、注意力和工作记忆，并将提供 对视觉信息处理模型的重要限制，以支持更有针对性的诊断和 临床环境中的干预措施。

Summary/Abstract Active vision requires encoding and remembering relevant information based on current task goals. Classic accounts posit that sensory encoding, attentional selection and working memory are mediated by persistent changes in the firing rates, or the gain, of visually responsive neurons that have a fixed tuning profile (termed “pure” or “fixed-selectivity” neurons). The focus on gain modulations in fixed-selectivity neurons has revealed a great deal about the basic mechanisms of attention and memory. However, it is becoming increasingly clear that dynamic changes in task demands may require more flexible coding schemes. For example, holding information in working memory in the same format as the stimulus-evoked response may lead to interference with new sensory inputs. Similarly, flexibly encoding sensory representations to complete one task – say a simple choice between two motor responses – might require a reconfiguration of the representation if another stimulus- response mapping suddenly becomes relevant. Finally, sensory codes must be flexible in the sense that early in processing they should form high-dimensional representations to represent as much information as possible about the current state of the world. Later in processing, when a decision or motor response needs to be made, the code should collapse to only represent the smaller subset of relevant choices. All of these computations are more naturally accomplished via the operation of neurons that have flexible tuning for both sensory features and for task demands (termed “mixed-selectivity” neurons). Based on these considerations, we hypothesize that flexible behaviors are supported by mixed-selectivity neurons that “rotate” high-dimensional neural codes to become robust to interference or to sub-serve other changes in task demands. We will use modelling, psychophysics, and functional magnetic resonance imaging (fMRI) to test predictions about how mixed-selectivity should modulate large-scale activation patterns that are measured non-invasively in human subjects. Collectively, this work will challenge traditional theories of sensory encoding, attention, and working memory that are based on the notion of fixed-selectivity, and will provide important constraints on models of visual information processing to support more targeted diagnoses and interventions in clinical settings.