CRCNS: Neural Basis of Inductive Bias

生物智能的一个标志是能够从一组非常小的经验中学习， 在新的环境中选择适当的行为。这种能力源于归纳偏见，指的是 用于将先前观察结果推广到新数据的假设。在机器学习中，归纳偏差是 这对于从少量的例子中有效学习至关重要，就像动物经常做的那样。在 在神经科学中，归纳偏见大多被认为是由进化形成的，以产生一种先天的 电感偏置虽然智能体的归纳偏差应该适应不断变化的任务需求，但如何 大脑发展归纳偏见，用它来指导学习，并使其适应任务统计，仍然很差 明白这个项目的总体目标是研究归纳的神经和计算基础 通过训练非人类灵长类动物进行新的学习任务，描述选择行为， 和神经元活动，并开发神经网络学习的计算模型。我们 神经核学习的理论框架可以精确预测行为和神经活动 这将在拟议的研究中进行实验测试。我们将用一种新的 设计的“串扰”任务，旨在表征感应偏差，并驱动受试者的 通过明确定义和可变的任务统计进行适应（目标1）。在这项任务中，受试者通过 对由多个特征组成的刺激进行分类的经验。在一组试验中，某些特征 是该类别的差异信息。接下来，我们将记录同步尖峰活动， 背外侧前额叶皮层和背侧纹状体的神经元在任务执行过程中（目的2）。在 同时，我们将开发和完善自适应学习的算法和人工神经网络模型 (Aim 3）生成对行为和神经活动的可测试预测。这些研究的结果将 影响范式用于研究学习和泛化的计算和神经基础， 人类和动物。

A hallmark of biological intelligence is the ability to learn from a remarkably small set of experiences to select appropriate behaviors in novel contexts. This ability arises from inductive bias, referring to the assumptions used in generalizing prior observations to novel data. In machine learning, inductive bias is essential for efficient learning from a small number of examples as animals frequently do. In neuroscience, inductive bias has been mostly considered as shaped by evolution to produce an innate inductive bias. Although an agent’s inductive bias should be adaptive to changing task demands, how brains develop inductive bias, use it to guide learning, and adapt it to task statistics, remain poorly understood. The overall goal of this project is to study the neural and computational bases of inductive bias adaptation by training non-human primates on a novel learning task, characterizing choice behavior and neuronal activity, and developing computational models of learning in neural networks. Our theoretical framework of neural kernel learning makes precise predictions for behavior and neural activity that will be experimentally tested in the proposed studies. We will characterize behavior in a newly designed “crosstalk” task, which is designed to characterize inductive bias and to drive a subject’s adaptation through well-defined and variable task statistics (Aim 1). In this task, the subject learns through experience to categorize stimuli composed of multiple features. Within a block of trials, certain features are differentially informative of the category. Next, we will record simultaneous spiking activity from neurons in dorsolateral prefrontal cortex and the dorsal striatum during task performance (Aim 2). In parallel, we will develop and refine algorithmic and artificial neural network models of adaptive learning (Aim 3) to generate testable predictions for behavior and neural activity. Results from these studies will impact paradigms used to study the computational and neural bases of learning and generalization in humans and animals.