Thursday, September 17, 2015

On the automatic process for training AI

To give a simple metaphor, we could think our AI model as a student.
A student is given exercises to prepare for the exams. They learn to solve these exercises by using some methods and they memorize those methods, not exercises themselves. The more exercises a student does, the more his methods are refined. Thus, his accuracy is improved. Our AI is similar.

Current static AI models is not as good as human brain. So there comes a point (after we train our AI with large volume of images), those methods that AI memorizes cannot be refined anymore. If we continue feed the AI more training images, it might learn new methods without retain old methods that it has learnt, and its performance could be worse. A student also cannot study all exercises in the world.

The complexity of methods that the AI can learn depends on the size of its model. An ambitious solution for above problem could be the dynamic AI model, which can increase its own size automatically as needed. This is hard and not guaranteed to be done correctly for now. Therefore, extending AI model size to capture more training images should be done manually by human engineer when it's possible.

Eventually, we will have extra large volumes of labeled images. Then we may take the approach of building an image search engine. This engine would take input images and provide similar results that can be used for image classification.

No comments:

Post a Comment