From Automation to Machine Learning
Artificial intelligence (AI) or machine learning (ML) has become a buzzword that we heard and see everyday everywhere. However, many people simply take machine learning as the same as automation. However, in fact they are very different from the first principle.
Let’s start from my two articles (Yiu and Yau, 2006 and Yiu, 2008) published in 2006 and 2008 on the 4 generations of intelligent buildings, and explain the fundamental differences between automation and machine learning. It would be so important that it makes machine learning a paradigm shift from traditional automation. After that, we can start discussing how to use Artificial Neural Network (ANN) to train a computer to make predictions on housing prices in another article.
The 4 generations of intelligent buildings are (1) Automated Building Model in the 1980s, (2) Responsive Building Model in the 1990s, (3) Effective Building Model in the 2000s, and (4) Learning Model since 2006. In the 1st generation, building facilities can perform designated functions automatically such as opening the door whenever it senses someone is coming. Then the 2nd generation, building facilities can make responses to the environment, for example, air conditioner’s temperature can adapt to the indoor and outdoor temperature and humidity. In the 3rd generation, it is often called Expert Systems, which possess all the experts’ knowledge and make predictions automatically to improve the efficiency and effectiveness of the building facilities. However, all the previous 3 generations of intelligence rely on the knowledge and instructions of human beings to be fed into the systems. The systems do not possess any learning capabilities to improve their performance themselves.
As raised by Penrose (2005, p.18) a learning algorithm does not specify the rules of operation in advance, ‘but instead there is a procedure laid down for the way that the system is to “learn” and to improve its performance according to its “experience”.’
In other words, automation is by a top-down approach, while machine learning takes a bottom-up approach. Automation requires hard-wired instructions which human beings know the answer. But machine learning can learn by itself, and may acquire knowledge even when human beings have not known the answer.
This forms the major difference between automation and machine learning. Just as Johnson (2004, p.150) mentions the purpose of learning is to ‘make rational prediction so as to overcome obstacles.’ As shown in the framework of an intelligent system below, whenever our decision leads to actions to achieve our goals, our senses would collect information and feedback to our brain to cross-check with our prediction, we learn by repeating the process and become our knowledge and experience. By means of categorization and analysis together with our instincts and beliefs, we can make better and better decisions and actions.
Based on this paradigm shift from automation to machine learning, nowadays computers and intelligent systems can learn how to improve themselves without relying on the knowledge of human beings. As the systems can learn even faster than human beings, it is logically possible that one day the intelligent systems can be smarter than us. That is why Harari (2017) worries that one day intelligent systems would control human beings.
Unlike the traditional automation model, machine learning requires a training period and a continuous learning process to update the knowledge, because the systems have to learn by this action-feedback-improved action loop. Interestingly, machine learning may acquire something very different from our knowledge and thus may make very different predictions and decisions, as their learning process and speed are different from us. In other words, from human beings’ perspectives, intelligent systems may make ‘wrong’ predictions and decisions.
Since 2006, machine learning has been developing very fast, and nowadays we can find applications of machine learning everywhere every day. For example, Facebook can recognize your friends’ image in your photo posts and tag with name automatically, but in fact, you have never told the systems who they are. In the real estate industry, machine learning has been exploited to help predict property prices. There have been lots of research and attempts to use Artificial Neural Network to predict housing prices. Let’s discuss in the 2nd article of this ML series.
Harari, Yuval Noah (2017) Homo Deus: A Brief History of Tomorrow.
Yiu, C.Y. and Yau, Y. (2006) A learning model of intelligent home, Facilities 24(9/10), 365–375. DOI:10.1108/02632770610677646
Yiu, C.Y. (2008) Intelligent building maintenance A novel discipline, Journal Of Building Appraisal, 2008, v. 3 n. 4, p. 305–317. DOI: http://dx.doi.org/10.1057/jba.2008.9