Source | Zhihu: Deep Blue Academy
Artificial intelligence is becoming increasingly popular today, with many terms constantly surrounding us: artificial intelligence, machine learning, deep learning, etc. Many people have a vague understanding of the meanings of these high-frequency terms and their relationships.
To help everyone better understand artificial intelligence, this article explains these terms in the simplest language, clarifying their relationships, hoping to assist beginners.
Figure 1 Application of Artificial Intelligence
1. Artificial Intelligence: From Concept to Prosperity
In 1956, several computer scientists gathered at the Dartmouth Conference and proposed the concept of “artificial intelligence”, dreaming of using the newly emerged computers to construct complex machines with essential characteristics similar to human intelligence. Since then, artificial intelligence has lingered in people’s minds and has slowly incubated in research laboratories. In the following decades, artificial intelligence has experienced extreme fluctuations, sometimes being hailed as a bright future for human civilization and sometimes dismissed as the fantasies of technological madmen. Until 2012, both voices still existed simultaneously.
After 2012, thanks to the increase in data volume, the improvement in computing power, and the emergence of new algorithms (deep learning) in machine learning, artificial intelligence began to explode. According to LinkedIn’s recent “Global AI Talent Report”, as of the first quarter of 2017, the number of technical talents in the global AI field based on the LinkedIn platform exceeded 1.9 million, with a domestic talent gap in artificial intelligence reaching over 5 million.
The research fields of artificial intelligence are also continuously expanding, with Figure 2 showing various branches of AI research, including expert systems, machine learning, evolutionary computation, fuzzy logic, computer vision, natural language processing, recommendation systems, etc.
However, current research work is mainly focused on weak artificial intelligence, which is likely to achieve significant breakthroughs soon. The artificial intelligence depicted in movies mostly refers to strong artificial intelligence, which is currently difficult to realize in the real world (artificial intelligence is typically divided into weak and strong AI; the former allows machines to have observational and perceptual abilities, achieving a certain level of understanding and reasoning, while strong AI enables machines to acquire adaptive capabilities and solve problems they have never encountered).
Weak AI has the potential for breakthroughs; how is it achieved, and where does “intelligence” come from? This is mainly attributed to a method of achieving artificial intelligence—machine learning.
2. Machine Learning: A Method to Achieve Artificial Intelligence
The most basic approach of machine learning is to use algorithms to analyze data, learn from it, and then make decisions and predictions about events in the real world. Unlike traditional software programs that are hard coded to solve specific tasks, machine learning is trained with large amounts of data, learning how to complete tasks through various algorithms.
For example, when we browse an online store, we often see product recommendations. This is because the store identifies which products you are genuinely interested in and likely to purchase based on your past shopping records and lengthy wish lists. Such decision models can help the store provide suggestions to customers and encourage product consumption.
Machine learning directly originates from the early field of artificial intelligence; traditional algorithms include decision trees, clustering, Bayesian classification, support vector machines, EM, Adaboost, etc. From the perspective of learning methods, machine learning algorithms can be divided into supervised learning (such as classification problems), unsupervised learning (such as clustering problems), semi-supervised learning, ensemble learning, deep learning, and reinforcement learning.
Traditional machine learning algorithms have reached commercial requirements or specific commercial levels in areas like fingerprint recognition, Haar-based face detection, and HoG feature-based object detection, but every step forward is exceptionally challenging, until the emergence of deep learning algorithms.
3. Deep Learning: A Technology to Achieve Machine Learning
Deep learning is not an independent learning method; it also employs supervised and unsupervised learning methods to train deep neural networks. However, due to rapid advancements in this field in recent years, some unique learning techniques have been proposed (such as residual networks), leading more people to view it as a separate learning method.
Initially, deep learning utilized deep neural networks to solve feature representation learning processes. Deep neural networks are not a completely new concept; they can be roughly understood as neural network structures with multiple hidden layers. To improve the training effectiveness of deep neural networks, adjustments have been made to aspects like neuron connection methods and activation functions. Many ideas had existed in earlier years, but due to insufficient training data and backward computing capabilities, the final results were unsatisfactory.
Deep learning has dramatically accomplished various tasks, making it seem that all machine-assisted functions are possible. Self-driving cars, preventive healthcare, and even better movie recommendations are just around the corner or about to be realized.
4. Differences and Connections Among the Three
Machine learning is a method to achieve artificial intelligence, while deep learning is a technology to achieve machine learning. We can use the simplest method—a Venn diagram—to visualize their relationships.
Figure 3 Diagram of the Relationships Among the Three
Currently, there is a common misconception in the industry that “deep learning may eventually eliminate all other machine learning algorithms.” This misconception arises mainly because deep learning is currently applied far more than traditional machine learning methods in fields like computer vision and natural language processing, and the media has exaggeratedly reported on deep learning.
Deep learning, as the hottest machine learning method today, does not mean it is the endpoint of machine learning. At least the following issues currently exist:
1. Deep learning models require a large amount of training data to exhibit miraculous effects, but in real life, small sample problems often arise, and deep learning methods cannot be applied; traditional machine learning methods can handle such situations;
2. In some areas, simple traditional machine learning methods can solve problems well, and there is no need to use complex deep learning methods;
3. The idea of deep learning is inspired by the human brain, but it is not a simulation of the human brain. For example, after showing a bicycle to a three or four-year-old child, they can still recognize it as a bicycle when they see another one that looks completely different, meaning that the human learning process often does not require large-scale training data, while current deep learning methods clearly do not simulate the human brain.
Deep learning leader Yoshua Bengio once said something particularly good in response to a similar question on Quora, which I will quote here to address the above issues:
Science is NOT a battle, it is a collaboration. We all build on each other’s ideas. Science is an act of love, not war. Love for the beauty in the world that surrounds us and love to share and build something together. That makes science a highly satisfying activity, emotionally speaking!
The general meaning of this passage is that science is not war but collaboration; the development of any discipline has never been a one-way street; it is about learning from each other, borrowing from each other, drawing on strengths, and continuously progressing while standing on the shoulders of giants. The study of machine learning is the same; cutthroat competition is a cult, while openness and inclusiveness are the right path.
Reflecting on the development of machine learning since 2000, Bengio’s words resonate deeply. Entering the 21st century, the development history of machine learning can be simply summarized into three research hotspots: manifold learning from 2000 to 2006, sparse learning from 2006 to 2011, and deep learning from 2012 to the present. Which machine learning algorithm will become the hotspot in the future? Feel free to discuss in the comments section!
ABOUT
关于我们
深蓝学院是专注于人工智能的在线教育平台，已有数万名伙伴在深蓝学院平台学习，很多都来自于国内外知名院校，比如清华、北大等。