| 2020-01-27 19:22 |
Lead: The next revolution in the field of machine learning began to sprout yet?
Lei Feng network AI Technology Review by: neural networks are excluded from academia, the scientific community true to his computer artificial intelligence, predictive modeling and a variety of deep learning tasks are taking on everything, but just ten years time. This decade we have witnessed a dropout, SGD + momentum, residuals access, a large outbreak of pre-training and so deep learning technology, academic research to embrace witnessed the depth of learning, also heard the sound of deficiencies depth study of the question of, many people feel the depth of the non-learning "revival" of expectations.
reddit recently appeared on a post, the friends who gather together to talk about new developments in machine learning methods in addition to the neural network, and can fit a variety of models, not just concepts and models to improve the neural network. Perhaps these are not popular now in the research on the seeds of the next revolution in machine learning, neural network research scholars like ten years ago ushered in the dawn of the revolution as in the cold.
The following Lei Feng network AI Technology Review compiled a number of users have substance answer.
Gaussian process Gaussian Processes
Gaussian process is not a new technology recent invention, but it is a significant development in recent years, especially in terms of implementation and expansion of technology. Compared to the neural network, Gaussian process characteristics are:
-
More intuitive, interpretability better (laugh saying that "Gaussian process is a set of smoother"), there are more mathematical tool to describe its behavior;
-
Very efficient, requires minimal sample and computing resources you can learn;
-
Can be easily integrated prior knowledge, intuition after setting with a set of parameters, it is likely no training you can get a good prediction;
-
Naturally in line with Bayes' rule.
Research Gaussian process's also incorporates many valuable mathematical discovery, linear algebra, probability, harmonic analysis involves so on. GPytorch is a Gaussian process model to understand the most outstanding field a good place.
The main deficiency in computing Gaussian process, training and reasoning process generally need to calculate the determinant and trace, or from a large matrix solver system, storage requirements increase as the square of the length of the column, and the time complexity degree of O (n ^ 3). Advances in recent years, mainly from the more efficient method of approximate calculation or algorithm (such as KISS-GP, SKI, LOVE etc.)
基因算法&演化策略 Genetic Algorithms & Evolution Strategies
离散演化训练是用基因算法配置网络结构，然后让得到的模型学习。它的一个动机来自于，在复杂环境中为稀疏的回报归因是非常困难的，所以不如完全抛弃梯度，转而采用计算更高效的演化策略，反倒可以在模型设计和参数搜索方面获得更大的灵活度、取得更好的结果。其中采用的和大自然中的生物演化类似的「随机突变+方向性选择」的做法也规避了当前的强化学习中的一些问题。
谷歌大脑的研究员 David Ha 在这个领域做了许多研究，他也撰写了易懂的介绍博客、配上了生动有趣的动图，可以点击 这里 详细阅读。
因果推理 Causal Inference
Judea Pearl 是这条路线的代表人物，Yoshua Bengio 等学者也加入了探索当中，并且在想办法把它和现代机器学习结合到一起。提到这个方向的这位网友表示，他的一位做数据科学的好朋友的工作内容里就已经越来越多地涉及到因果推理，而且他认为因果推理带来的革命在日后看来可能会比深度学习革命的影响更为深远。接着他做了详细的解释：
首先要知道的是贝叶斯网络。1980 年代的时候，Judea Pearl 正在探索各种能对人工智能领域的发展起到帮助的路线，贝叶斯网络的发明就有一部分功劳是他的。但是贝叶斯网络也有一些限制，它能在低维空间高效地捕捉联合概率分布，但是说到底它也只能回答一些可观测的问题。比如，给定了一些消费者的属性，基于其它消费者的行为，预测这些消费者在未来六个月内不再继续使用某个服务的概率。
但值得研究的问题还有很多。理想情况下，如果你要采取一些行动干预某个系统，你会希望能够了解系统会发生哪些变化（由于行动干预）。沿着刚才消费者的例子说，如果把他们加入一个为了提高顾客忠实度和互动程度而设计的自动发送促销邮件的清单里，他们不再继续使用服务的概率会如何变化？这时候你需要思考的就是，给定了已知的消费者信息，也指定了要不要用某种方式干预之后，得到的结果会如何变化。这是因果关系运动的一个方面，目前看来 Rubin 和 Imbens 是这一方面的权威人物。详细了解可以参考这篇文献综述 http://proceedings.mlr.press/v67/gutierrez17a.html 。
另一方面，你相当于在估计这个量 E[Y|X, do(T)]，这里的 Y 是得到的结果、X 是观察到的条件、T 是你要采取的措施。有没有更通用的理解因果关系的方式呢？我很欣赏 Pearl 的分解方式，他展示了超越贝叶斯网络之外的处理方式，而且可以把过程表示为一个因果图模型。他的思路是，图模型里的箭头方向可以编码因果关系的方向，而干扰某个系统就可以看作是破坏图里的几个边。 继续消费者的例子，也许你手里有一些曾经收到促销邮件的消费者的数据，但他们的情况不一样，你想知道其它情况的消费者看到促销邮件以后的反应如何；而且，你手里现有的数据肯定不是双盲的（因为是发现消费者有不再使用的意向之后才给他们发送促销邮件）。所以，本来会根据客户表现出的某些信号来决定是否给他们发送促销邮件，但真的发了以后又会干扰这些信号；而且，对于不同来源的用户，引发“要发送邮件”决定的具体信号又会各自不同…… 所以要怎么办呢？在图上做这些分析计算就可以帮助回答这些问题，或者，最起码也可以知道哪些问题是可以回答的、哪些是回答不了的，以及想要回答现在回答不了的问题的话还需要哪些信息。
Judea Pearl 在 2017 年写的《Book of Why》是一本非常棒的入门书籍，这个领域的每个研究者都应该读一读这本书，它很容易读懂，只不过对于喜欢听音频书的人来说，里面谈图元素的部分还是最好用看的。如果你想做更深入的研究，Pearl 2009 年的学术专著《Causality》数学性更强，也更严谨，不过书里的习题和例子都很少，所以想要理解里面的所有内容还很需要花一些功夫。
《A Meta-Transfer Objective for Learning to Disentangle Causal Mechanisms》（https://arxiv.org/abs/1901.10912）论文讲解了因果推理和深度学习之间的联系，最起码也要读读开头的几小节。更大更全的综述可以读读 Pearl 的《Theoretical Impediments to Machine Learning With Seven Sparks from the Causal Revolution》（https://arxiv.org/abs/1801.04016），不过它实际涵盖的内容和《Book of Why》有很多重合，先读读这篇论文再决定要不要买书来读也挺好的。
反向强化学习 Inverse Reinforcement Learning
反向强化学习采用了和传统强化学习相同的基础设定，然后做相反的事。在强化学习里，给定一个回报函数，让模型找到会得到最大回报的策略；在反向强化学习里，给定一个策略，然后模型找到可以被这个策略最大化的回报函数。
它的关键在于从对行为的观察中学习，即便你可能无法访问回报函数，或者无法模仿特定的执行器的行为。反向强化学习已经成功训练出了和人的行为很类似的玩第一人称射击游戏的 AI。
Reverse reinforcement learning has a major open question is how to learn from than optimal presentation. The current system can well imitate human beings, that they reflect the failure mode and also similar to humans. Want to learn from humans, but to get a higher performance than humans, in theory, it appears to be feasible, but doing it is very difficult. You can refer to a recent review, https://arxiv.org/abs/1806.06877 .
Automatic machine learning AutoML
Since this field is very important, too cool. It can be seen as a decision tree to help you decide what kind of data processing pipeline after a given data set is the best. It is useful, will occupy more places in the whole field of machine learning. Google currently has for business users AutoML service, but obviously this is not for his own client programming capabilities of.
The most popular automatic machine learning algorithm is sklearn-Auto ( https://automl.github.io/auto-sklearn/master/ ), as well as its further improvement based on Mosaic. AutoML there are some games, the top few algorithms have very high reference value.
In addition to the above, users are also referred to the optimal transmission theory Optimal Transport Theory, symbolic regression Symbolic Regression, spiking neural networks Spiking neural networks, stochastic optimization Stochastic optimization, there are some ways to get inspiration from biology and so on. If you know some interesting, machine learning methods have significant progress in recent years, we welcome comments and discussion.
