Infants' self-generated visual statistics support object and category learning

项目摘要 人类视觉对象识别在其具有挑战性的识别单个对象的能力方面是显着的 它可以帮助我们在不同的环境中快速识别，甚至可以快速识别成千上万个日常类别的新实例。 尽管人们对这些成熟期的过程了解很多，但对它们的发展却知之甚少 特别是对于常见的日常对象和支持鲁棒对象识别的体验 和分类。这一差距是至关重要的，因为对象识别和分类支持早期单词 学习，解决物理问题，以及后来学习拼写和数学符号。这 研究项目的重点是1岁幼儿的视觉物体学习，这是一个发展时期， 视觉物体识别的显著进步的前端和具有多种风险因素的儿童 开始落后于规范的发展轨迹。该方法侧重于真实的的属性- 世界视觉体验，支持学习在具有挑战性的视觉环境中识别单个物体， 将这种学习推广到同一类别的成员。该方法使用头戴式眼动仪捕捉 来自100名17至22个月大的婴儿的视野图像，因为他们自发地互动和玩耍， 对象通过与物体的积极互动，婴儿产生了自己的视觉数据包用于学习。 与物体感知相关的多种视觉特性将通过算法进行测量和量化。幼儿情绪 对主动参与的对象和来自同一类别的新对象的识别将被测量， 具有挑战性的基准上下文，包括杂乱、遮挡和不同的视图。类别泛化将 在一个名称概括任务中进行测量。先进的统计和机器学习将决定视觉 支持婴儿物体识别和分类的自我生成经验的属性。研究 将首次提供幼儿与物体积极互动的自然视觉统计特征 以及潜在的变革性证据，证明人类在视觉对象方面的能力的发展基础 分类并不在于经验与许多不同的情况下，一个单一的类别，标准 假设，但在积极的视觉经验与个别对象。此外，有发育风险的婴儿 语言延迟和自闭症显示了早期物体名称学习的中断，最近与 对物体的视觉学习中断。该项目包括对1999年至2000年的高危婴儿进行初步分析， 为下一步的长期研究计划做准备。

PROJECT SUMMARY Human visual object recognition is remarkable in its ability to recognize individual objects in challenging circumstances and to rapidly recognize even novel instances of tens of thousands of everyday categories. Although a great deal is known about these processes at maturity, very little is known about their development especially with respect to common everyday objects and the experiences that support robust object recognition and categorization. This gap is critical because object recognition and categorization support early word learning, physical problem solving, and the later learning of orthographies and mathematical symbols. This research projects focuses on visual object learning in 1 year old toddlers, a developmental period that at the front end of marked advances in visual object recognition and a period in which children with multiple risk factors begin to fall behind the normative developmental trajectory. The approach focuses on the properties of real- world visual experiences that support learning to recognize individual objects in challenging visual contexts and generalizing that learning to same category members. The method uses head-mounted eye-trackers to capture field-of-view images from 100 infants 17 to 22 months of age as they spontaneously interact and play with objects. Through active interactions with objects infants generates their own packets of visual data for learning. Multiple visual properties relevant to object perception will be algorithmically measured and quantified. Toddlers’ recognition of the actively-engaged object and a novel object from the same category will be measured in challenging benchmark contexts including clutter, occlusion, and different views. Category generalization will be measured in a name generalization task. Advanced statistics and machine learning will determine the visual properties of self-generated experiences that support infants object recognition and categorization. The research will provide the first characterization of the natural visual statistics of toddlers’ active interactions with objects and potentially transformative evidence that the developmental foundation for human prowess in visual object categorization lies not in experiences with many different instances of a single category, the standard assumption, but in active visual experiences with individual objects. Moreover, infants at risk for Developmental Language Delay and Autism show disruptions in early object name learning that have been recently linked to disruptions in visual learning about objects. The project includes preliminary analyses of infants at risk in preparation for the next step in the long-term research program.