When the identity of R & D personnel experienced the PC Internet and mobile Internet development process of the two, is now facing a new wave of wave of AI, although identity has not become personally engaged in the technical work of the business managers. But as the subversion of the Internet has experienced various industries / enterprises / enabling, AI remodeling industry / enterprise must also be the trend for many non-technical background or insensitive technology enterprise managers, AI / machine learning simply means a noun a trend we are chasing a packaging-to-business ....... But for a person who still love technology, in-depth understanding of the value lies in how the enterprise is no longer a packaging / product combination with AI, but from analytic business scene, abstract business scene selection, data collection, algorithms .... .. the landing part of the enterprise business and AI were combined to form a competitive first-mover advantage.
I had previously seen a lot of AI-related books and information, good and bad, be spoon-feeding education. All kinds of history, full of concepts in the mind, a kind of feeling foggy. But this is also good for spoon-feeding education, is when faced with a number of excellent books or articles, to take advantage of its clear hierarchy of these scattered knowledge to sort out the association. For example, know the historical development of artificial intelligence, symbolism, connectionism, behaviorism, knowing a core technology of artificial intelligence, machine learning only, but the depth of learning and is a subset of machine learning, neural networks realize the depth of learning the base model, image recognition, speech recognition, natural language processing application depth study of machine learning ..... because now almost become synonymous with AI (core), it became the focus of everyone learning areas, very fortunate to find "Scikit-learn and TensorFlow machine learning practical Guide (photocopy edition)" this book, the first chapter in the book let his machine learning, there has been a clear and systematic understanding. Other chapters and case depth and the code is also very good, with practical and enforceable.
Read other books about artificial intelligence
"Smart Age" in 2016 published a book, when artificial intelligence in its infancy, ready to chase the wave of this wave. So remember a passage from the book:. "This is the best of times, worst of times to homeopathy, or contrarian down, that our own choices."
"Artificial Intelligence Revolution" feeling pretty good a history of artificial intelligence, technology, science books characters, although the evaluation said very shallow, but each book has its own position should it - popular science books. At the same time learned from this book a great British drama "Black Mirror" , chasing a few quarters, considered a windfall it. ^ _ ^
"Science ultimate: Talk artificial intelligence" is a popular science book, but I feel this book is reflected in Zhekesiwei, more difficult to understand.
"Collective intelligence program" 15 years of publication, watercress score 8.9 high scores, but only a handful of readers. I think AI and hot era of big data, an obscure title will let you miss a good book, want to "machine learning" to understand and practice the "statistical computing," a friend recommended reading, which scenario is very close to life (such as product preferences recommendation, forecast prices, dates match, financial scene ......), the code is appended can do ...... but understanding this book accompanying the knowledge points more than the estimated need reading practice times, and a lot of learning to expand (¯ (●●) ¯)
"Introduction to Artificial Intelligence" is similar to an academic textbook books, more formal, embraces a wide range, suitable for entry-learning artificial intelligence guide.
: "Fission seconds artificial intelligence to understand the basic course" book of good organization and context, such as learning to master the mathematical basis of what artificial intelligence; the most fiery of machine learning, artificial neural networks, deep learning is like; examples of neural network; artificial intelligence beyond what depth learning, artificial intelligence, machine learning is not all; scenarios of artificial intelligence. This progressive layers of context to explain yourself have a more clear understanding of the framework of artificial intelligence. In addition to the book author and author of "artificial intelligence revolution", writing is very good, both the technology and humanism.
Leisure Books:
"A Brief History of the Future" "A Brief History Today" Yuval • Hela Li's works, a lot of conflicts and crises in the future to explore the human and machine intelligence may face.
"Artificial Intelligence" "AI · Future" Kai-fu Lee teacher's books, easy to understand. Few technical aspects of things, and thinking mainly about the relationship between individuals, society and artificial intelligence.
Attached "Scikit-Learn and TensorFlow Machine Learning Practical Guide (photocopy edition)" The first chapter exercises
How to define the machine learning?
Machine learning is programmed to make computer science (and art) to learn from the data.
Machine learning is a computer with the ability to learn, without explicit programming.
The computer program uses the experience of learning tasks T E, P performance, if for task performance P T, with the growing experience of E and P also will increase, it is called machine learning.
Machine learning can solve any problem, please List four?
The problem requires a lot of manual adjustment or need to have a long string of rules to resolve: Machine learning can simplify the code and improve performance.
Complicate matters, traditional methods difficult to solve: Use good machine learning techniques can find a solution.
Fluctuations in the environment: machine learning algorithm can adapt to the new data.
Discover unexpected relationships / trends from complex issues and large amounts of data (data mining) in a better way to deal with the problem.
What is the training set with a label?
In supervised learning, the training data used to train the algorithm contains the answer, called a tag.
Please list two common-supervised learning task?
Classification, such as spam filters is to train with a lot of mail with classified samples, but also for new mail filter to classify;
Return (forecast target value), for example, give some characteristic (mileage, vehicle age, brand, etc.), to predict the price of a car.
Pointed out that four common non-supervisory tasks?
Clustering, using a large number of user data is assumed blog visitor, detecting packet similar visitors.
Visualization and dimension reduction, the data for a large number of complex and algorithm tagged, the algorithm 2D or 3D image data is output.
Anomaly detection, detecting an abnormality such as a credit transfer to prevent fraud, detect manufacturing defects, or outliers removed automatically from the training set before training data.
Association rule learning, mining large amounts of data to find interesting relationships between attributes.
If a robot can walk on a variety of unknown terrain, what would you machine learning algorithm?
Reinforcement Learning: learning system referred to herein as agent, can be observed on the environment, selection and implementation of the action, rewarded (negative reward is punishment). Then it must own to learn what is the best method (strategy), in order to obtain maximum long-term rewards. Policy determines the action agent in a given situation should be taken.
要对你的顾客进行分组，你会采用哪类算法？
在不知如何定义顾客组的情况下采用非监督学习的聚类算法；如果你知道你想要什么类别的用户组，采用监督学习的分类算法。
垃圾邮件检测是监督学习问题，还是非监督学习问题？
监督学习问题
什么是在线学习系统？
在线学习是用数据持续的进行训练，可以一次一个或者一次几个实例（小批量）。每个学习步骤都很快且廉价，所以系统可以动态地学习到达的数据。
什么是核外学习？
在线学习算法也可以当机器的内存存不下大量数据时，用来训练系统（成为核外学习）。算法加载部分的数据，用这些数据进行训练，重复这个过程，直到用所有数据都进行了训练。
什么学习算法是用相似度做预测？
基于实例学习的机器学习算法使用记忆来学习案例数据。然后使用相似度测量推广到新的例子。
基于模型学习算法的参数和学习算法的超参数的区别是什么？
基于模型学习算法有一个或多个模型参数，这些参数决定了它将如何预测给定的新实例（例如，线性模型的斜率）。此学习算法试图找到这些参数的最优值，这样模型就可以很好地推广到新的实例。学习算法的超参数是是其本身的参数，而不是模型的。
基于模型学习的算法搜寻的是什么？最成功的策略是什么？基于模型学习如何预测？
基于模型的学习算法为模型参数寻找最优值，使模型能够很好地推广到新的例子中。我们可以定义一个实用函数（或拟合函数）用来测量模型是否够好，或者定义一个代价函数来测量模型有多差。对于线性回归问题，人们一般是用代价函数测量线性模型的预测值和训练样本的距离差，目标是使距离差最小。
机器学习的四个主要挑战是什么？
训练数据不足。机器学习需要大量的数据，才能让多数机器学习算法正常工作。即便对于非常简单的问题，一般也需要数千的样本，对于复杂的问题，比如图像或语音识别，你可能需要数百万的样本。（也可重复使用部分存在的模型）。
没有代表性的训练数据。
低质量的数据。如果训练集中的错误、异常值和噪声太多，系统检测出潜在规律的难度就会变大，性能就会降低。
不相关的特征。避免进来的是垃圾，出去的也是垃圾。
如果模型在训练集上表现好，但推广到新实例表现差，问题是什么？给出三个可能的解决方案？
出现了训练集的过拟合。过拟合发生在相对少量的训练数据，噪声较多，模型过于复杂的情况。解决方案：
简化模型，可以选择一个参数更少的模型（比如使用线性模型，而不是高阶多项式模型）、减少训练数据的属性数、或者限制一下模型）；
收集更多的训练数据；
减少训练数据的噪声（比如修改数据错误和去除异常值）。
什么是测试集，为什么要用它？
在将算法模型应用在产品中前，使用测试集来进行评估模型推广误差率。
验证集的目的是什么？
用训练集和多个超参数训练多个模型，选择在验证集上有最佳性能的模型和超参数。
如果用测试集调节超参数，会发生什么？
如果在测试集上多次测量了推广误差率，调整了模型和超参数，以使模型最适合这个集合。这意味着模型对新数据的性能不会高。
什么是交叉验证，为什么它比验证集好？
为了避免“浪费”过多训练数据在验证集上，通常的办法是使用交叉验证：训练集分成互补的子集，每个模型使用不同的子集训练，再用剩下的子集验证。一旦确定模型类型和超参数，最终的模型使用这些超参数和全部的训练集进行训练，用测试集得到推广误差率。