深度学习的能与不能

[摘要] 深度学习技术的应用日渐广泛，在语音、图像、文本处理、搜索引擎、广告推荐等领域都取得了巨大的成功。认为深度学习自身具有盲点，无法解决全部的机器学习问题，并指出了深度学习的优缺点，为深度学习的使用者提供了一定的理论指导。最后，还展望了深度学习的未来发展趋势。

[关键词] 深度学习；傻瓜型学习算法；专家型学习算法；白箱算法；黑箱算法；相关性；因果性

[Abstract] Deep learning is now widely used and has achieved huge successes in many fields, such as speech, image, natural language processing, search engine and advertising recommendation. In this paper, considering the achilles' heel of deep learning, we propose that deep learning cannot solve all problems in machine learning discipline. Moreover, the advantages and the disadvantages of deep learning are illustrated, and some suggestions for users are proposed. Finally, the future trend of deep learning is pointed out.

[Keywords] deep learning; fool-type learning algorithm; expert-type learning algorithm; white box algorithm; black box algorithm; relevance; causality