Depth of an article with you into unsupervised learning from the basic concept to the four implement gamelink

The depth of | an article take you into unsupervised learning: from basic concept to four implementation models (of the Sohu) – Culurciello’s blog technology from Eugenio Author: Eugenio Culurciello machine of the heart: Li Yazhou, Wu Jing compiled in June this year the Purdue University associate professor Eugenio Culurciello wrote an article about unsupervised learning an overview of the article. In addition to the basic concepts, this paper also introduces four kinds of implementation models of unsupervised learning: cluster learning, automatic encoder, generative model, PredNet. A few days ago, Professor Culurciello based on the recent development of unsupervised learning this article has been updated and adjusted, the heart of the machine was compiled. The papers mentioned in this article can be downloaded by reading the original text. Note: Liu Diwei (translator), Liu Xiangyu (Revised) two teachers of the June version compiled and released to the CSDN geek headlines, this article compiled before the article borrowed two teachers of translation (some adjustment), if there is not acceptable, please contact the machine of the heart, thank you! Unsupervised learning is the Holy Grail of deep learning, the goal is to establish a general system compatible with small data sets for training, even with very little data. Today, deep learning models are often trained on large supervised datasets. The so-called supervised data set, that is, each data has a corresponding label. For example, the popular ImageNet data set, there are one million artificially labeled images. A total of 1000 classes, each class has 1000 images. It takes a lot of effort and a lot of time to create such data sets. Now imagine creating a dataset with 1M classes. Imagine, for each frame of the video data set with the 100M data frame. The task is immeasurable. Now, think about how you learned when you were a child. Yes, there will be someone who will guide you, and your parents will tell you it’s a cat, but they won’t tell you it’s a cat every minute of the rest of your life! The same is true of supervised learning today: I tell you time and time again, what is a cat, perhaps up to 1 million times. Then your deep learning model learns. Ideally, we would like to have a model that is very similar to our brains. Only a small number of tags can be understood in this multi class world. Here said the class, mainly refers to the object class, action class, the environment class, object class, etc.. The main objective of the basic concept of unsupervised learning is to train a model (called "recognition" or "coding") for other tasks. Coding features can often be used in a classification task: for example, training on ImageNet results in very good results, which is very close to the monitoring model. So far, the monitoring model is always better than unsupervised pre training model. The main reason is the special n of the monitoring model相关的主题文章: