Throughout the ages, mankind has been exploring the limits of technology. With the outbreak of information technology in the 21st century data science and artificial intelligence technology to usher in the spring of their own, especially in depth study based on artificial intelligence techniques can be said to shine, far better than humans in many fields, such as face and such recognition technology have also landed, even pessimists believe that the era of intelligent machines, if one day the machine with independent thinking, humanity will face extinction. However, the depth of learning it so omnipotent? the answer is negative! Depth study is the use of deep neural network technology, having been able to transcend the human image recognition and so on, but it still has many aspects that can not be completed, the paper cited a number of areas deep learning are not currently implemented, hoping to help you open mind, a better understanding of the depth of learning.
Depth study "can"
Pa is both a science, but also the king of game
In general, most people can see 300 documents per year, while IBM's Watson system in 10 minutes where you can read 20 million documents, it is obvious ability to learn the depth of learning is far greater than mankind is full of "learning Pa . " At the same time, in a game field, whether it is or go dota2, deep learning has enough ability to crush humanity. So deep learning is both a school bully, the game is king.
versatile
Chess, writing poetry, composing, art paintings ?????? 2016, Alpha Dog victory Shishi, 2017 Microsoft wheatgrass published his first collection of poems, followed by turning to music ????? already depth study gradually become well versed in literature, multi-talented all-powerful king.
From the depth of learning outcomes achieved point of view, it seems to have been all-powerful, in many aspects than human.
Depth study "can not"
Algorithm output is unstable, vulnerable to attack
In the field of image recognition, we may change only in one image pixel values of a point, then the output will be a huge change, which is caused by unstable output of the algorithm, this subtle change seems insignificant in humans, for algorithm model is really different. Not only in the image field, natural language processing also have this problem. In the Q & A system, in the original text have to add some random simple words, the ability to understand the model is greatly reduced. This problem occurs not only in the depth of learning, traditional machine learning more vulnerable to attack.
模型复杂度高，难以纠错或调试
在2016年阿法狗与李世石的大战中，李世石赢了一局。在李世石的78手后，阿法狗的胜率便直线下降。如果可以投降的话，那么在李世石的第78手后，阿法狗应该会选择投降，而并不会针对这一手进行相应的改进。此外，在深度学习进行翻译时，不管是给模型什么数据输入，都会有一个有意义的输出。此前的谷歌翻译曾遇到过这样的问题，在翻译结果有明显错误的时候，翻译部门的工程师也很难去对模型修改，可见深度学习模型的复杂。
层级复合程度高，参数不透明
在图像识别领域，我们在模型的中间层中尽力去抓取图像的特征。在第一层的卷积层计算后，我们对结果进行可视化，可以很容易看出结果与原图像有很大相似性。然后，随着层数的加深，对中间其他层的可视化，我们完全不能看出中间层所代表的意义。主要原因在于感受野的复合，而且每层的卷积核也会产生复合，加上一些模型会有自己特有的复合，如inception模块的复合，残差的复合，让我们难以从中间层的可视化中看到模型具体运行的结果。
对数据依赖性强，模型增量性差
深度学习是端到端结构，灵活性非常低。我们将单个图像拼接在一起，人类很容易识别的内容，深度学习确无能为力，可见其迁移能力较差。在“语义标注”和“关系检测”这类问题中，人类可以通过完成一个任务中的多个子任务，并将子任务整合的方式解决问题，而对于深度学习来说，多个子任务与一个总任务是完全不同的两个任务，需要不同的模型去解决问题。在数据量较小的情况下，模型拟合能力较差。
专注直观感知类问题，对开放性问题无能为力
我们小时候都曾学习过关于乌鸦喝水的故事。乌鸦在面对半瓶水，而自己的嘴够不着水时，会往瓶子里丢入石子，使得水面上升从而喝到水。此外，乌鸦在无法拨开坚果时，它会把坚果丢在马路上，让来往的车辆碾压从而迟到果实，在此过程中，乌鸦能够通过观察人行道的情况学会判断车辆是否会行驶以保障自己的安全。而鹦鹉也有自己的智能，在听过人类重复说过的话后，鹦鹉能够很好地模仿人类说话。深度学习只能做到鹦鹉的智能，而做不到乌鸦的智能，可见其泛化能力之低。此外，深度学习也难以理解图像背后的寓意。当一幅图中出现奥巴马与一群大象时，深度学习仅仅能辨认图中是一个男人与一群大象，显然图作者却是想透过图片暗喻美国的两党之争，一般来说，大象喻指美国民主党。
机器偏见难以避免，人类知识难以有效监督
这可能是目前深度学习面临的最大问题。数据是深度学习的基础，而数据的可靠程度决定了模型的可靠程度。微软层开发聊天机器人Tay，模仿年轻网民的语言模式。但是试用24小时后便被引入歧途，成为偏激的种族主义者，甚至发出了“希特勒无罪”的消息。原因在于年轻的网民本身的语料库并不是纯净的，是人就会有偏见，这种偏见在网络中尤其严重，这样便导致了Tay用来训练的数据带有偏见，并使得Tay误入歧途，而人类知识的监督很难有效采用，这就无法避免机器的偏见。另一个例子，美国法院用以评估犯罪风险的算法COMPAS，也被证明对黑人造成了系统性歧视。机器偏见无法消除，日后可能会给人类带来严重的后果。
总结
不可否认，深度学习可以在特定领域超过人类，有很好的效果，但它并非万能。某种意义上说，它离智能还差很远。目前，对深度学习的泛化性与可解释性的呼声越来越高。2017年7月，国务院在《新一代人工智能发展规划》中提出“实现具备高可解释性，强泛化能力的人工智能”。或许下一代人工智能技术还是在深度学习基础之上展开，但是希望新的技术能够很好地解决现在深度学习的不能，更好地造福人类！