Learning higher level representations and invariant transformations

我们能否设计出一种人工智能系统，一旦陷入未知世界，就像新生儿第一次睁开眼睛一样，能够完全从它接收到的海量原始感觉信息中形成高水平的概念？这样的系统应该能够在这个不断变化的信息流中发现深刻的规律性和稳定的结构，以便它最终可能形成自己对空间、时间、物体(或多或少稳定的实体)运动的主观概念。本研究旨在探索新的基本原则，以指导具有这种能力的下一代自主学习系统的构建。我们的方法将建立在计算机实现的基础上，这些计算机实现能够不断增长“理解层”，一个接一个地叠加。每一个这样的层都应该通过利用前一层未能捕捉到的规则来构建对前一层提取的表示的稍微更高级别的“理解”。我们打算专注于一类特殊的规则，称为“不变变换”，我们希望自动发现和学习它。层层提取更高层次、更多含义的表示，有望成为人工学习系统中接近人类水平的表现的关键因素(从而推动机器学习、模式识别和数据挖掘领域的最先进水平)。预计它还将生产出计算机化的系统，能够根据过去的观测数据进行更准确的预测。这将在许多领域有直接的技术应用，在这些领域，卓越的预测系统非常有用：从工程到金融，从商业决策到医学。

Can we design an artificial intelligence system that, once plunged in an unknown world, like a newborn baby opening his eyes for the first time, would be able to form high level concepts exclusively from the enormous quantity of raw sensory information that it receives? Such a system should be able to discover deep regularities and stable structures in this ever changing information stream, so that it may eventually form its own subjective notions of space, time, of things (more or less stable entities) moving around. This research program aims at exploring new fundamental principles to guide the building of next generation autonomous learning systems with this kind of abilities. Our approach will be based on computer implementations capable of growing "layers of understanding", stacked one upon the other. Each such layer shall be building a slightly higher level "understanding" of the representation extracted by the previous one, by exploiting regularities that the previous layer failed to capture. We intend to focus on a particular class of regularities called "invariant transformations", that we want to automatically discover and learn. The layer by layer extraction of ever higher level, ever more meaning carrying representations, is expected to be a key element for opening the door to near human-level performance in artificial learning systems (thus advancing the state-of-the-art in the fields of machine learning, pattern recognition and data mining). It is also expected to produce computerized systems capable of far more accurate predictions from past observed data. This would have direct technological applications in a wide range of fields where superior prediction systems are immensely useful: from engineering to finance, from business decision making to medicine.