1. Introduction to Deep Learning

1.1 AI map

① As shown in the figure below, the X-axis is different modes. The earliest one is semiotics, then probability model and machine learning. The Y-axis is what we want to do. Perception is when I understand what it is, reason to form my own knowledge, and then make plans.

② Perception is similar to I can see a screen in front of me. Reasoning is based on what I see and imagine what will happen in the future. Based on the phenomena and data I see, I form my own knowledge. After knowing all the knowledge, I can make long-term plans. What to do in the future.

① Natural language processing currently still focuses on perception. Things that a person can respond to in a few seconds belong to the scope of perception, even if it is like translating Chinese into English and English into Chinese.

② Computer vision can do some reasoning in pictures.

③ There are symbols in natural language processing, so there is semiotics, and probabilistic models and machine learning can also be used. Computer vision is faced with pictures, and pictures are filled with pixels. It is difficult to explain pixels using semiotics, so most computer vision uses probabilistic models and machine learning to explain.

④ Deep learning is a method in computer vision, and it has other application methods.

1.2 IMAGENET data set

① Deep learning first made a major breakthrough in image classification.

② IMAGENET is a relatively large image classification data set, as shown in the figure below. It includes pictures of a thousand categories of natural objects, and it has about one million pictures.

1.3 Image classification error rate

① It can be seen that starting from 2012, the error rate of image classification began to decline significantly. 2012 was the beginning of the introduction of deep learning into image classification.

② In 2017, almost all teams were able to achieve an error rate within 5%, which basically reached human accuracy in image recognition. It can be said that deep learning has done very well in image classification.