Member-only story
Transfer Learning for Image Classification — (2) Pre-trained Image Models
Image classification is the task to recognize an image. It is also called image recognition. Computer scientists have been innovative in extracting meaning from images. Its history is fascinating, though most people don’t know much about it. For this reason, I am going to tell you the stories of innovation in this chapter. You will see how researchers tried to solve the challenges. You will learn some prominent pre-trained models. Then I will show you how to use both PyTorch and Tensorflow to load a pre-trained image classification model to predict an unknown image. My code is available in this PyTorch notebook and this Tensorflow notebook.
(A) The Genesis of ImageNet
Artificial Intelligence (AI) took off in the early 2000s. Back at that time, most AI researchers were developing better algorithms to recognize images. The research community faced an unprecedented change — the lack of a large collection of image data. Not only that, all the images should be annotated. How can we build a very rich image database so that the categories of images can represent all the objects comprehensively? Professor Fei-Fei Li at Stanford University took on the challenge. She turned her attention to building the image database. She and a small team had an ambitious goal — to map out the entire world of objects.
How to map out the entire world of objects? It needs a comprehensive classification of words and concepts. In 2007 she met with Princeton professor Christiane Fellbaum, a major WordNet contributor. WordNet is a large database for the semantic relations between nouns, verbs, adjectives, and adverbs. It can be seen as a large digital dictionary and thesaurus. Wouldn’t it be great if images are attached to the words in WordNet? That is the genesis of ImageNet. Professor Fei-Fei Li and the team attached images to each category of WordNet. She called the image database ImageNet. WordNet was the fruit of Linguistics and has been applied widely in Natural Language Processing (NLP). The two disciplines cross-pollinated to produce the new fruit — ImageNet.
