课程: Self-Supervised Machine Learning

Constructing self-supervised tasks

课程: Self-Supervised Machine Learning

开始一个月试用 购买团队方案

Constructing self-supervised tasks

“

- [Instructor] Self-supervised tasks have to be constructed from the data available. How do you do that? Well, let's say you have time series data, historical data. You can set it up in such a way that you predict the future from the distant past. You can also predict the future from the recent past. It depends on what you're trying to learn. You can also switch things around. You can mask out the past and try to predict the past from the data that you have available from the present. Or let's say you have an image of a kind. You can predict the top of the image from the bottom or the bottom from the top. By masking portions out you predict the occluded from the visible. That's also an image technique. For example, in natural language processing, let's say you have the entire sentence available in the raw form. And then you mask out a word and try and predict the future from the past, to predict the word in the future from the words you have from the past sequence. Or you can use the same sentence, and mask out the word early in the sequence, and use the later words of the sequence to predict the missing word, to predict the past from the present. You have a bunch of raw unlabeled images. You simply cut off, and make available the bottom of the image, and ask the model to predict the top from the bottom. Or you cut off the bottom of the image, make available the top, and ask the model to predict the bottom from the top. Another way to set up a self-supervised task is to predict the occluded from the visible, mask out a portion of the image, and use the rest of the image to predict what lies in the masked-out portion. Now, a good question would be, these tasks are great, but they seem like, well, just fun. How do these tasks actually help? This takes us to the core idea behind self-supervised learning. A model that is trained to predict hidden features from observed features can learn generalized representations of the underlying data. And these generalized representations will then be useful in other downstream tasks.

内容