KDI: Learning of Objects and Object Classes in Visual Cortex

学习分类和识别物体的能力是人类和高等动物视觉系统的一个关键特征。然而，这些强大的视觉能力背后的表征是如何组织和获得的，在很大程度上仍然是未知的。研究者和他的同事们通过计算、生理和心理物理方法的结合，解决了视觉皮层如何学习表征和识别新物体和物体类别的问题。对警觉的猴子进行的生理实验依赖于多电极记录，研究人员为此开发了一套适当的数据挖掘技术工具包，部分基于他们自己在学习和分类算法方面的工作。特别是，研究人员承担了一个多学科的研究项目，包括四个相互作用的组成部分：i)颞下皮层神经元的计算建模，扩展了他们以前在IT中单个对象表示的工作；Ii)皮质生理学，在清醒的、行为良好的猴子身上使用多个电极，训练它们在新的刺激类别上进行类间和类内分类任务；Iii)处理多电极数据的新数据挖掘技术，包括分类和学习技术；iv)视觉心理物理学，包括人类和猴子的功能磁共振成像研究，允许将猴子生理学的发现与人类大脑中的物体学习联系起来。理解人类大脑中的学习意味着理解智力的核心。这不仅是科学中仍然存在的基本挑战之一，而且也是一个即使是微小的进步也会对理解神经疾病和紊乱，以及对计算和机器智能的未来产生重大影响的领域。然而，尽管在过去的一二十年里取得了巨大的进步，科学界还不知道大脑皮层的各个区域有什么作用以及如何起作用。因为理解大脑——我们所知道的最复杂的系统——是一项巨大的努力，目前的项目侧重于理解大脑皮层的一部分，这涉及到每个人日常生活中一个关键而非常困难的任务——即使主观上很容易：学习分类和识别视觉对象，如面孔或汽车。了解脑细胞是如何表现物体的将是神经科学的一个重大突破，也是最终设计出能够达到人类表现的机器的一个重大突破。更重要的是，在物体识别这一特定问题上的任何重大进展都将对KDI项目的目标产生重大影响，因为它将为理解大脑和机器中学习和智能的更广泛问题打开大门。

Poggio9872936The ability to learn to categorize and recognize objects is a key feature of the visual system of humans and higher animals. Yet, how the representation underlying these powerful visual abilities is organized and acquired is still largely unknown. The investigator and his colleagues tackle the problem of how the visual cortex learns to represent and recognize novel objects and object classes through a combination of computational, physiological and psychophysical approaches. The physiological experiments on alert monkeys rely on multielectrode recordings for which the investigators develop a tool kit of appropriate data mining techniques, based in part on their own work on learning and classification algorithms. In particular, the investigators undertake a multi-disciplinary research project consisting of four interacting components: i) Computational modeling of inferotemporal (IT) cortical neurons, extending their previous work on representations of single objects in IT; ii) cortical physiology using multiple electrodes in awake, behaving monkeys trained on between- and within-class classification tasks on novel classes of stimuli; iii) new data mining techniques for processing multiple electrode data, including classification and learning techniques; iv) visual psychophysics including fMRI studies in humans and monkeys, allowing to relate the findings from monkey physiology to object learning in the human brain.Understanding learning in the human brain means understanding the very core of intelligence. Not only is this one of the remaining fundamental challenges in science but it is also one area where even small steps forward will have significant implications for understanding neurological diseases and disorders, and also for the future of computing and machine intelligence. However, despite enormous progress in the last decade or two, science does not yet know what various areas of the cortex do and how. Because understanding the brain, the most complex system we know, is a huge endeavor, the present project focuses on understanding a part of cortex, involved in a key and very difficult task in everybody's daily life -- even if subjectively very easy: learning to categorize and recognize visual objects such as faces or cars. Understanding how brain cells come to represent objects will be a major breakthrough for neuroscience and also for eventually designing machines capable of achieving human-like performance. More importantly, any significant progress in the specific problem of object recognition will have a major impact on the goals of the KDI program, because it will open the door to understanding broader issues of learning and intelligence in brains and machines.