Learning Algorithms for Deep Architecture Systems (Algorithmes d'Apprentissage pour Systèmes à Architecture Profonde)

在人工智能领域成立50多年后，我们仍然对人类的抽象思维能力感到困惑。当面对在机器中实现这种复杂思维的艰巨任务时，人们注意到人类思维可以操纵的概念的丰富性和多样性。人类可以查看编码场景视频的3D像素阵列，并从中产生几个不同的高级（抽象的）具有不同结构的表示：它包含的对象集，这些对象之间的关系网络，场景语义的文本描述等。我们如何在计算机中重新创建，这种人类在这些丰富的结构之间平滑过渡并保留它们所代表的真实的世界的核心语义的能力？ 这项研究计划将试图回答这个问题，通过从人类大脑中的深层计算组织中汲取灵感。它提出研究可以适应人工神经网络行为的算法，组织在一个深层架构中，这样它就可以将具有某种任意结构的观察作为输入，并对其进行转换，以便输出具有可能不同结构的所需目标。深度学习系统是针对具有有限结构复杂性的数据的问题而开发的，因此改善了此类问题的最新技术水平。该研究计划旨在为目前没有深度学习解决方案的结构化数据问题带来相同的改进。这种自适应的预测系统将在许多领域中得到应用，例如计算机视觉和自然语言处理。更重要的是，这项研究的结果将揭示从模拟经验中构建能够操纵高级抽象概念的自主智能系统的原理，并缩小人工智能和人类智能之间的差距。

More than 50 years after the founding of the field of artificial intelligence, we are still baffled by the capacity of humans for abstract thought. When faced with the daunting task of implementing such complex thinking in a machine, one notices the richness and variety of concepts the human mind can manipulate. Humans can look at the 3D array of pixels encoding the video of a scene, and produce from it several different high-level (abstract) representations with varying structures: the set of objects it contains, the network of relationships between these objects, a textual description of the semantics of the scene, etc. How could we recreate, in a computer, this ability that humans have to smoothly transition between these rich structures and preserve the core semantics of the real world they represent? This research program will attempt to answer this question, by taking inspiration from the deeply layered organization of computations in the human brain. It proposes to investigate algorithms that can adapt the behavior of an artificial neural network, organized in a deep architecture, so that it can take as input an observation with some arbitrary structure and transform it so as to output a desired target with a possibly different structure. Deep learning systems have been developed for problems with data of limited structural complexity and have thus improved the state-of-the-art on such problems. This research program will aim to bring the same improvements to problems with structured data that do not have deep learning solutions at present. Such adaptive, predictive systems will find application in a number of fields, such as computer vision and natural language processing. More fundamentally, the results of this research will shed light on principles for building, from simulated experience, autonomous intelligent systems capable of manipulating high-level abstract concepts, and shrink the gap between the state-of-the-art in artificial intelligence and human intelligence.