Exploding models | 6900 likes of AI learning roadmap on Medium, allowing you to quickly get started with machine learning

Source: http://blog.csdn.net/wemedia/details.html?id=43739

Translation | AI Tech Base Camp (rgznai100)
Participation | thinkdeeper


Part 1: Why Machine Learning is So Important

AI Tech Base Camp Editor's Note:

This article With a total of over 6900 likes on Medium, it is extremely popular on Medium. The biggest reason for its popularity is that the author can explain the various principles included in machine learning clearly and thoroughly in relatively simple words. And this is simply a benefit package for the following three groups of people:

technicians who want to quickly improve machine learning capabilities;
non-technical people who want to have a preliminary understanding of machine learning and are willing to contact related concepts;
anyone who thinks about how machines think those who are interested.

Although the article will also discuss the basic concepts of probability, statistics, programming, linear algebra and calculus, but under the author's simple language description, even without relevant background knowledge, it will not be difficult to understand.

If you want to quickly understand the concepts of machine learning in 3 hours, and do not know where to find valuable instructive articles, I recommend you to read this author's series of articles. AI Technology Base Camp will update Part 2, 3, 4, 5 and the resource list of this series one after another.

Part 1: Why Machine Learning Matters. Overview of Artificial Intelligence and Machine Learning - Past, Present, Future.
https://medium.com/machine-learning-for-humans/why-machine-learning-matters-6164faf1df12

Part 2.1: Supervised Learning. Linear regression, loss functions, overfitting, gradient descent.
https://medium.com/machine-learning-for-humans/supervised-learning-740383a2feab

Part 2.2: Supervised Learning II. Two classification methods: logistic regression and SVMs.
https://medium.com/machine-learning-for-humans/supervised-learning-2-5c1c23f3560d

Part 2.3: Supervised Learning III.. Nonparametric learning: k-nearest neighbors, decision trees, random forests. And introduce cross-validation, how to adjust parameters and model fusion.
https://medium.com/machine-learning-for-humans/supervised-learning-3-b1551b9c4930

Part 3: Unsupervised Learning. Clustering: k-means, hierarchical clustering. Dimensionality reduction: Principal Component Analysis (PCA), Singular Value Decomposition (SVD).
https://medium.com/machine-learning-for-humans/unsupervised-learning-f45587588294

Part 4: Neural Networks. How deep learning works, along with Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNNs), and practical applications.
https://medium.com/machine-learning-for-humans/neural-networks-deep-learning-cdad8aeae49b

Part 5: Reinforcement Learning. Introduce the Markov decision process. Q-learning, policy learning, deep reinforcement learning. Value learning problem.
https://medium.com/machine-learning-for-humans/reinforcement-learning-6eacf258b265

Addendum: Best Machine Learning Resources List of Machine Learning Course Resources.
https://medium.com/machine-learning-for-humans/how-to-learn-machine-learning-24d53bb64aa1

This is the first article in a series, the author tells you why machine learning is in the way of storytelling so important.

This is reminiscent of the debate between Xiao Zha and Musk about "whether AI will surpass human beings" not long ago. Xiao Zha said that AI only helps people improve efficiency and allow people to focus on more important work; while Musk made it public on twitter Pi Xiaozha "understood too shallowly and did not see a longer-term threat".

The point of this article is the same as that of Musk. The author is full of worries about the future, and it is this worry that he proposes:

Although predicting the future is a bit unrealistic, one thing is certain: 2017 is the best year for understanding how machines think. Good time to start.

In his view, people must go deep into machines to understand the internal details of the world, to figure out what machines really want, what their biases and weaknesses are.

To understand these issues, the author emphasizes that "machines must be studied as human beings"; just like the study of psychology and neuroscience, which explores topics such as how humans learn, make decisions, and feel. From this level, this is not just a technical problem, but requires a combination of logic, psychology, philosophy, linguistics and other disciplines to truly understand the secrets behind the black box.

Text:

Artificial intelligence is more powerful than any other innovation of this century, and it will shape our future more powerfully.

In this world ruled by technology, anyone who doesn't understand it will soon find that he is left behind... He will be stunned by the upcoming magical world.

Although the ebb and flow of artificial intelligence in the past dynasties has been rotten by major information articles, it has long been a cliché, but please be patient, I still want to use my own way to sort it out.

After experiencing several AI stagnation and development cycles in the past four decades, the explosion of data and the improvement of computing power have finally broken the bottleneck of artificial intelligence.
 
In 2015, Google trained a conversational bot (AI) that not only communicates with people as technical support, but also discusses ethical issues, expresses opinions and answers general fact-based questions.
 

Vinyals & Le, 2017

In the same year, DeepMind developed a program that surpassed human performance in 49 Atari games using only pixels and game scores as input.

Soon after, DeepMind once again achieved self-transcendence, releasing a new cutting-edge game algorithm called A3C.

At the same time, AlphaGo beat the world's top Go player, the second major victory in a human-dominated game after two decades of machines conquering chess.

Many Go masters do not understand why a machine can master this ancient Chinese strategy game, how can a machine master the essence and complexity of it, and defeat humans in 10¹⁷⁰ possible layouts. You know, the number of atoms in the universe is 10⁸⁰. This is incredible.
 

The picture shows professional Go player Lee Sedol replaying the game after a failed match with AlphaGo.

However , this is not over yet.

In March 2017, OpenAI developed a robot in a language it invented, effectively achieving its goal. Before long, Facebook successfully trained bots that could negotiate and even lie.

On August 11, 2017, in the 1v1 match of the multiplayer online game Dota 2, OpenAI defeated the world's top game players, which shocked a lot of people.


  To watch the whole game, please go to YouTube, Dendi (human) vs. OpenAI (bot)

In addition to these shocking counterattacks, artificial intelligence has also begun to permeate our daily lives.

For example, in the following two pictures, you only need to point the camera at the menu, and the selected picture can be automatically translated into English through the Google Translate APP.



The next time you stay in a hotel, call the front desk to have toothpaste delivered to your room. Don't be surprised when you open the door and see a small housekeeping delivery robot in front of you.



Various segments have also begun to use artificial intelligence for intelligent upgrades.

In healthcare today, AI can be used to design evidence-based treatment plans for cancer patients, to analyze medical test results in real-time, and for drug discovery.

( Editor 's Note: You can refer to an article compiled by AI Technology Base Camp before, "An article about the nine footholds of artificial intelligence in the medical industry, so that you can understand AI better") In 5 chapters, the core machine learning concepts behind these techniques are explored in depth, and you are taught to use tools to build similar applications.)

Below I give a picture of a knowledge tree, from which you can see that machine learning mainly includes supervised learning, without There are three aspects of supervised learning and reinforcement learning.


 
Artificial intelligence is primarily the study of robots that can perceive the world around them, make plans, and make decisions to achieve their goals. The required foundations include mathematics, logic, philosophy, probability, linguistics, neuroscience and decision theory. Many fields fall under AI, such as computer vision, robotics, machine learning, and natural language processing.
 
Machine learning is a subfield of artificial intelligence. Its goal is to let the computer learn by itself. Machine learning algorithms enable them to identify patterns in data to build models, rather than making predictions with predefined rules.


The Big Bang of Intelligence

In fact , the examples discussed above, although surprising, are all weak artificial intelligence (ANI), but it does solve certain tasks efficiently.

We've been doing some ground-level advancement work for Strong Artificial Intelligence (AGI). AGI is defined as one that can perform all kinds of complex tasks required by humans, including learning, planning, and decision-making under uncertainty, communicating in the same language as humans, telling jokes, directing stock trades, or programming yourself.

It is especially important to know how to program it yourself. Once we create a self-improving AI, it will improve itself in a recursive and recursive fashion. This heralds that, at some point in the future, we will enter an era of intelligent explosion, where

superintelligent machines can surpass the intellectual activity of any human being. From this, it can be deduced that designing this super-intelligent machine is one of human intellectual activities, and the super-intelligent machine is greater than the human's intellectual activities, so the super-intelligent machine can design a better machine.

As a result, there will be an intellectual explosion, far exceeding the big explosion of human intelligence.

It can be said that the first superintelligent machine was the last human invention.

The graph below is called a singularity. The term is borrowed from the gravitational singularity that occurs at the center of a black hole, a one-dimensional point of infinite density. Here, the laws of physics we are accustomed to understand no longer apply.



We have absolutely no idea what's going on inside a black hole, because no light can escape the black hole.

Likewise, we have no way of predicting what will happen after we unlock AI's self-improvement capabilities.

Just like the guinea pig used in human experiments, it never knows what humans are doing to itself, it can only superficially understand that humans are helping them get more cheese.

Recently, the Future of Humanity Institute published a survey that surveyed AI researchers' predictions for the realization of the era of AGI.

Researchers believe that in 45 years, AI will have a 50% chance of winning over humans. Of course, some say it's been longer, some say it's been a few years.



This is a picture from the 2005 book Near the Singularity. Now, in 2017, how many more self-righteous pictures can we have on the wall?

Is the emergence of super artificial intelligence a good thing or a bad thing for human beings?

Instead of making bad predictions, let's talk about a very practical question: How can we designate an AI to do what we want him to do in a human-friendly way?

To achieve this goal, we need to understand machine learning, not to apply machine learning, but to really understand.


Machine Learning - The Only Way to Avoid Machine Attacks

So I've always been certain of one thing: 2017 is the time to start understanding how machines think.

This is not a slap in the head to understand and learn AI with the help of philosophical abstractions or other figurative metaphors. Rather, we must dig deep and understand the internal details of the machine world—what they “want”, their underlying biases and weaknesses—just as we study psychology and neuroscience to understand how humans learn, make decisions , behave and perceive this way.

Questions about the complexity and high stakes of AI will be increasingly raised in the coming years. Many questions need to be answered:

How do we deal with AI's apparent tendency to systemic bias in existing datasets?
What should the world's most powerful technologists think of the fundamental disagreement about the potential risks and benefits of AI?
In a world without jobs, what is the human sense of purpose?

Machine learning is at the heart of artificial general intelligence and will transform every industry and have a huge impact on our daily lives. That's why we think it pays to learn about machine learning. At least on a conceptual level, I write these series of articles more suitable for getting started, and there is no feeling that people are shut out a thousand miles away.

So with all that said, how to read this series of articles?

You don't have to read the entire series, but here are three suggestions, depending on your interests and timing.

1. T-type method. Read from beginning to end. and summarise each section. This way you will be more motivated to read, and less likely to be forgotten. To get to the bottom of what interests you most, we'll provide some references at the end of each section for you to expand on.

Reference: Feynman technique
address: https://mattyford.com/blog/2014/1/23/the-feynman-technique-model

2. Focus method. Go straight to the part that interests you most.

3. The 80/20 Rule. Just skim through it all, take notes on the main concept, and understand it.

For the 80/20 rule, you can refer to:
https://www.thebalance.com/pareto-s-principle-the-80-20-rule-2275148

Okay, that's all for today's opening chapter, please pay attention to AI Technology Base Camp WeChat for continuous updates.

About the authors of this series:



Vishal most recently worked at Upstart, a lending platform that leverages machine learning to price credit, automate the lending process, and acquire users. His research is based on how to think about entrepreneurship, applied cognitive science, moral philosophy and the ethics of artificial intelligence.

Samer is a MS in Computer Science and Engineering at UCSD and co-founder of Conigo Labs. Before graduating, he founded TableScribe, a business intelligence tool for small and medium businesses. He spent two years advising McKinsey's Fortune 100 companies. Samer studied Computer Science and Ethics, Political Science and Economics at Yale University.

The pair's goal is to solidify their understanding of artificial intelligence, machine learning, and how they fit together, and hopefully create something worth sharing in the process.

original address
https://medium.com/machine-learning-for-humans/why-machine-learning-matters-6164faf1df12