The role of long-term experience in the development and use of the visual number sense

每当你使用智能手机、上传内容到社交媒体、或与Netflix或Spotify互动时，你都在与神经网络互动：一种经过训练的机器学习算法，可以更好地识别和分类你和你朋友的面孔，你的电影和音乐偏好，一类神经网络--俗称“深度网络”--引起了心理学家和神经科学家的特别兴趣，因为这些网络的结构和行为特别类似于大脑处理视觉信息的方式。因此，深度网络可能会提供机器和人类智能之间缺失的环节，阐明支持大脑表示，处理和预测抽象信息的能力的生物机制。 深度网络在多大程度上代表了人类大脑实际上是如何代表世界的？我们通过视觉数字感来探索这个问题：所有人类从出生起就拥有的感知系统，在日常生活中快速估计数字时使用，例如猜测罐子里有多少弹珠。视觉数字感不仅展示了大脑如何表示高度抽象的信息，而且还与其他认知能力相关，包括我们在儿童和成人的正式学校教学数学中的表现。最近，研究人员设计了视觉数字感的深度网络，这些网络与人类的数字感知表现出惊人的一致性（例如，附近的数字更难区分，网络性能的改善与儿童在发育过程中的网络性能高度相似，等等）。但是，为了真正测试这些网络是否是人类数字感知的适当模型，我们需要从这些网络中获得原始的、尚未测试的预测，并检查这些相同的模式是否发生在人类观察者身上。在这项研究中，我们使用最先进的深度网络模型，首先推导并测试关于人类数字感知在儿童和成人中应如何表现的预测。例如，这些网络预测，当我们经历更多的真实世界场景时，视觉数字感就会出现并进行微调--碗里的苹果，走在街上的人，等等。因此，实验刺激越接近真实的场景，而不是人工场景，数字感知就应该越好。通过使用真实世界的场景（包括照片和3D软件中创建的场景），我们可以检查成人在估计与自然主义场景统计数据更一致的场景中的数字时是否表现出优势，以及这种优势是否解释了儿童数感改善的发展轨迹。另一方面，如果我们在现实世界的场景中没有观察到数字感知的任何优势，我们就会知道，虽然深度网络是机器数字感知的优秀模型，但它们并不能充分捕捉人类观察者视觉数字的出现、发展和使用。

Every time you use your smartphone, upload content to social media, or engage with Netflix or Spotify, you are interacting with a neural network: a machine-learning algorithm that has been trained to become better at recognizing and categorizing faces of you and your friends, your movie and music preferences, etc. One class of neural networks - colloquially known as "deep nets" - has been of particular interest to psychologists and neuroscientists, because the structure and behaviour of these networks in particular resembles how the brain processes visual information. Deep nets might therefore provide the missing link between machine and human intelligence, elucidating the biological mechanisms that support the brain's ability to represent, process, and predict abstract information. To what degree are deep nets representative of how human brains actually represent the world? We explore this question through the visual number sense: a perceptual system all humans have from birth that is used in everyday life when quickly estimating number, such as when guessing how many marbles are in a jar. The visual number sense not only showcases how the brain represents highly abstract information, but also correlates with other cognitive abilities, including how well we do in formal, school-taught mathematics as children and adults. Recently, researchers have designed deep nets of the visual number sense that have shown surprising alignment with human number perception (e.g., nearby numbers are harder to distinguish, the improvement in network performance is highly similar to that of children as they develop, etc.). But, to really test whether these networks are an appropriate model of human number perception, we need to derive original, still untested predictions from these networks and examine whether these same patterns occur in human observers. In this grant, we use state-of-the-art deep nets models to first derive and then test predictions about how human number perception should behave in both children and adults. For example, these networks predict that the visual number sense emerges and fine-tunes as we experience more real-world scenes - apples in a bowl, people walking on a street, etc. Therefore, the closer the experimental stimuli are to real, as opposed to artificial, scenes, the better number perception should be. By using real-world scenes (both photos and those created in 3D software), we can examine whether adults demonstrate advantages when estimating number in scenes that are more vs. less consistent with naturalistic scene statistics, and whether this advantage explains the developmental trajectory at which children's number sense improves. If, on the other hand, we fail to observe any advantages for number perception in real-world scenes, we would know that while deep nets are excellent models for machine number perception, they do not adequately capture the emergence, development, and use of visual number in human observers.