Uncovering the neurobiology of combined supervised and unsupervised learning

揭示监督和无监督学习相结合的神经生物学

• 批准号: RGPIN-2014-04947
• 负责人: Richards, Blake
• 金额: $ 2.11万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2015
• 资助国家: 加拿大
• 起止时间: 2015-01-01 至 2016-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=588111
• 关键词: Uncovering; neurobiology; combined; supervised; unsupervised

Mathematical models of learning tell us that there are two different ways to learn from experience. One way is to learn to associate a sensory input with a desired output. For example, when we learn to read as children we learn to make particular sounds when we see certain characters. This type of learning is called “supervised learning”. In contrast, another way to learn is to pick out patterns in our experiences. For example, people who have a lot of experience with wine learn to identify particular patterns and categories in smells and tastes. This type of learning is called “unsupervised learning”. Researchers in artificial intelligence have shown that when you combine supervised and unsupervised learning computers can learn more rapidly and generalize better. Intuitively, this makes sense: once a person has a great deal of experience with wine and knows about the patterns in smells and tastes, he or she can assign labels to these patterns, fruity, earthy, etc., which will aid learning about new wines in the future. Despite the theoretical importance of combined supervised and unsupervised learning, neuroscientists know very little about how it works in the brain. In fact, one of the only examples neuroscientists have of this combined learning is when young animals begin to process the spatial relationships between sights and sounds. Deep in an evolutionarily ancient part of our brains we combine visual signals and auditory signals to locate objects in the world around us. Our ability to do this is not hardwired genetically; rather, it is something that we learn when we are young. First, we learn about the spatial patterns that exist in our visual sensations, i.e. we use unsupervised learning to locate objects with our eyes. After that, we learn how to associate the sounds we hear to the spatial information that our eyes provide, i.e. we use supervised learning to locate objects with our ears. For example, experiments have shown that if you raise an animal with goggles on its face that shift everything they see to the left, then the auditory maps of space in these deep brain regions will also be shifted to the left to match the information coming from the eyes. Although we know that this ancient part of our brains combines supervised and unsupervised learning in this way, there is a great deal that we don't understand about how it accomplishes this. In particular, it is unclear how the brain ensures that our eyes teach our ears how to locate objects, rather than the other way around. Theoretically speaking, there is no reason why our auditory systems couldn't learn without visual inputs and then teach our visual systems how to locate objects; after all, congenitally blind people can accurately use their ears for spatial localization. So, what mechanism does the brain use to ensure that, when everything is normal, our eyes teach our ears? My laboratory's research will attempt to answer this long-standing question. To understand the complex interactions occurring within the brain we will use a combination of electrical recordings in live neurons, genetic tools for controlling neurons with light, and computer models. We will determine how visual and auditory inputs to these deep brain regions are altered at different ages of development, and we will use what we learn to answer the question of how visual inputs teach auditory inputs. This research will provide a critical and timely contribution to the study of learning in neural systems, helping to close a gap between our mathematical and neurobiological theories. In the future, this could help us develop new therapies for improving learning and enable brain-computer interfaces for linking our brains with artificial learning systems.

学习的数学模型告诉我们，从经验中学习有两种不同的方法。一种方法是学习将感官输入与期望的输出关联起来。例如，当我们小时候学习读书时，我们会在看到某些字符时学会发出特定的声音。这种类型的学习称为“监督学习”。相反，另一种学习方法是从我们的经历中找出模式。例如，对葡萄酒有丰富经验的人学会识别气味和味道的特定模式和类别。这种类型的学习称为“无监督学习”。人工智能研究人员表明，当将监督学习和无监督学习结合起来时，计算机可以学习得更快，概括能力也更好。直观上，这是有道理的：一旦一个人对葡萄酒有了丰富的经验，并且了解气味和味道的模式，他或她就可以为这些模式分配标签，果味、泥土味等，这将有助于将来学习新葡萄酒。 尽管监督学习和无监督学习相结合在理论上很重要，但神经科学家对其在大脑中的运作方式知之甚少。事实上，神经科学家关于这种组合学习的唯一例子之一是幼年动物开始处理视觉和声音之间的空间关系。在我们大脑进化古老的部分深处，我们结合视觉信号和听觉信号来定位我们周围世界的物体。我们这样做的能力并不是与生俱来的，而是与生俱来的。相反，这是我们年轻时学到的东西。首先，我们了解视觉感觉中存在的空间模式，即我们使用无监督学习来用眼睛定位物体。之后，我们学习如何将听到的声音与眼睛提供的空间信息相关联，即我们使用监督学习来用耳朵定位物体。例如，实验表明，如果你养一只脸上戴着护目镜的动物，它们看到的一切都会向左移动，那么这些大脑深处区域的空间听觉图也会向左移动，以匹配来自眼睛的信息。 尽管我们知道我们大脑的这个古老部分以这种方式结合了监督学习和无监督学习，但我们仍然不明白它是如何实现这一目标的。特别是，目前还不清楚大脑如何确保我们的眼睛教我们的耳朵如何定位物体，而不是相反。从理论上讲，我们的听觉系统没有理由不能在没有视觉输入的情况下学习，然后教我们的视觉系统如何定位物体；毕竟，先天性失明的人可以准确地利用耳朵进行空间定位。那么，大脑使用什么机制来确保，当一切正常时，我们的眼睛教导我们的耳朵呢？我的实验室的研究将试图回答这个长期存在的问题。为了了解大脑内发生的复杂相互作用，我们将结合使用活体神经元中的电记录、用光控制神经元的遗传工具和计算机模型。我们将确定这些深部大脑区域的视觉和听觉输入在不同的发育年龄是如何改变的，并且我们将利用我们学到的知识来回答视觉输入如何教导听觉输入的问题。这项研究将为神经系统学习的研究提供关键而及时的贡献，有助于缩小数学和神经生物学理论之间的差距。未来，这可以帮助我们开发出改善学习的新疗法，并使脑机接口能够将我们的大脑与人工学习系统连接起来。