Imagenet and Model Fitting
Robert H. Chen, Chelsea Chen · 2024
For a machine to be able to identify and classify something, it would first have to have a database of almost everything. ImageNet is a database of 100 million images and their annotations, including 62,000 images of cats alone. The accuracy of computer vision machines was demonstrated at the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) where GoogLeNet&s;s 22-layer, 9-inception module DCNN won with 94% accuracy defeating not only the competitor machines but also routinely performing better than human beings at image recognition. In artificial neural networks, the mapping function is constructed by training the network on a set of labeled data; that is, by induction from the abstracting of common characteristics found in the labeled training set data to form a model generalization for recognizing newly presented data by deduction, going from the model to a proper specific case. All AI models are subject to the overfitting (not generalizable) and underfitting (too generalizable) of input data, and there are many methods to achieve accurate generalizable by proper fitting.