Classifying Images and Detecting Objects
Neural networks have gone from being one curiousity in lists of classification methods to being the prime engine of a huge and very successful industry. This has happened in a very short time, less than a decade. The main reason is that, with enough training data and enough training ingenuity, neural networks produce very successful classification systems, much better than anyone has been able to produce with other methods. They are particularly good at classifying images. As Fig. 18.1 shows, the top-5 error rate on one (very large and very hard) image classification dataset has collapsed in quite a short period. The primary reason seems to be that the features that are being used by the classifier are themselves learned from data. The learning process seems to ensure that the features are useful for classification. It’s easy to see that it might do so; the news here is that it does.