SPP 1527: Autonomous Learning

近年来，计算学习研究在解决许多数据分析问题方面取得了巨大的成功。机器学习方法和统计学习理论已经成为各种工程、生命科学和自然科学学科中不可缺少的工具。然而，这些方法在很大程度上依赖于专家收集数据并以某种适当的格式表示它，决定合适的算法和超参数，以及决定内部表示的结构。这与学习应该导致更多自主性的意图相矛盾，它与我们在生物系统中观察到的学习形成了对比。这一优先方案的目的是发展自主学习系统的新基础。这需要新的概念和方法，超越现有的机器学习方法，走向自主探索未知环境并开发适当表示的系统。自主学习的核心方面是：(1)自主选择(超)参数、表征和学习特征；(2)自主收集数据，即探索和主动搜索以加速学习，而不是从静态数据集中学习；(3)自主发展适当的表征，包括层次结构；(4)刺激、内部表征和动作的增量提取。现有的机器学习和机器人学方法，特别是强化学习，为这项研究提供了一个起点。在此基础上，我们的目标是迈向真正自主学习系统的下一步。

In recent years, computational learning research was tremendously successful in solving many data analysis problems. The methods of machine learning and statistical learning theory have become essential tools in various engineering, life science and natural science disciplines. However, such methods depend to a large degree on an expert to collect the data and represent it in some appropriate format, to decide on a suitable algorithm and hyper-parameters, and to decide on the structure of internal representation. This contradicts the intention that learning should lead to more autonomy and it contrast to learning as we observe it in biological systems. The aim of this Priority Programme is to develop novel foundations of autonomously learning systems. This calls for new concepts and methods, which go beyond existing machine learning methods, towards systems that autonomously explore an unknown environment and develop appropriate representations. Core aspects of autonomous learning are: (1) the autonomous choice of (hyper-)parameters, representations and features for learning, (2) the autonomous collection of data, i.e., exploration and active search to accelerate learning instead of learning from static data sets, (3) the autonomous development of appropriate representations, including hierarchies, and (4) the incremental abstraction of stimuli, internal representations and actions. Existing machine learning and robotics methods, in particular reinforcement learning, provide a starting ground for this research. Based on this, we aim for the next step towards truly autonomously learning systems.