Autonomous Deep Learning for AI

点击翻译按钮获取中文摘要

There has been impressive progress in artificial intelligence (AI) in the last few years thanks to the development of systems based on deep learning, the modern instantiation of artificial neural networks. These advances have been mostly on perception tasks, using supervised learning (often from human-labeled data), and for cognitive tasks - called system 1 tasks - which are performed fast, in an unconscious way, and associated with intuition (i.e., not easy to verbalize). Much remains to be done on the side of actions, as well as cognitive tasks - called system 2 tasks - which are performed slowly and sequentially, in a conscious way, and associated with reasoning and expressed verbally. Actions are important because they allow the learner to intervene in its environment, thus to more easily uncover causal relationships. Uncovering causal explanations is important to generalize far from the typical situations covered by the training data (and humans are much better at this than machines, currently). Actions are also important because they allow the learner to actively seek knowledge, to purposely explore its environment to better understand it. Finally, actual applications of AI of course often require machines to act in complicated ways, e.g., learn a policy, and not just perform a classification. In this context, the proposed research combines the following five interconnected research objectives: (1) explore grounded language learning methods based on jointly learning language and about the world to which it refers (2) explore agent-learning methods to acquire knowledge (including about cause and effect) about the environment that we wish the learner to refer to with natural language (3) develop new learning theory that does not rely on the assumption of a same training and test distribution to one where only the causal mechanisms need to be preserved (4) design deep learning architectures and training frameworks that exploit such a theory in order to learn independent causal mechanisms which, when combined, can be used for building causal explanations and for planning, as well as to lift deep learning from system 1 to joint system 1 and system 2 abilities, and (5) train such systems on grounded language learning and world modeling tasks in virtual environments with a fast research cycle.