How can AI Behave Like Humans? Discuss "Machine Learning" and "Deep Learning"

Human wisdom is accumulations of experience. Learning from successes and failures, we will be able to make the best decisions based on our past memories when encountering similar incidents in the future.

The most commonly mentioned AI technologies are "machine learning (ML)" and "deep learning (DL)", and their relationship with AI can be expressed by the following concentric circles.

In brief, DL is part of ML, but both technologies are in the realm of artificial intelligence.

ML is the process of training a machine to learn by itself through large amounts of data and algorithms. From tons of data, the computer uses statistical analysis to find changing patterns of information, summarize features, and then conducts specific tasks. In this way, developers do not need to write programs repeatedly for different situations, as the computer can learn to make accurate judgments itself. For example, social media algorithms analyze pictures, articles, and videos that their users usually browse or like, and recommend contents to them according to the user's preferences.

As mentioned above, ML includes a large number of algorithms, such as Logistic Regression, Decision Tree, Random Forest, K-means Clustering, etc. These algorithms can be roughly divided into three categories: Supervised Learning, Unsupervised Learning, and Semi-supervised Learning, based on "whether the data has standard answers." 

Supervised learning gives the computer labeled data (conceptually similar to standard answers), from which it can collect features, train a model, and make predictions. For example, we input 5,000 photos of a dog and a cat into a computer, and then another photo is interpreted by the computer to confirm that it is a dog or a cat. During the learning process, the computer continuously corrects itself to achieve more accurate predictions until its error of judgement reaches an acceptable range.

Unsupervised learning involves training a model via unlabeled data. Without providing standard answers, the computer finds hidden rules in the data by itself. For example, if you input information of 10,000 company's customers, the computer will automatically place people with the same gender, similar age and other characteristics in a group. Semi-supervised learning, on the other hand, has some labeled data, but usually the amount of it is much less than the unlabeled one.


Let's move on to deep learning. Professor Hong-Yi Lee in the Department of Electrical Engineering at NTU once said, "Deep learning is a set of functions, and that's all." The algorithmic model of DL is structured with a neural-like network, which is a function set composed of a large number of neurons that do simple computations. When we input a dataset, the computer outputs another set of values, and then picks one of them as the best computed solution. DL applications include AlphaGo and autonomous driving technology, and have been integrated with medical services these years in order to create more advanced applications in the AI industry.