📌 Usage of techniques – In production tips
Throughout this book, you will come across “In production” tip boxes like the following one, highlighting real-world applications of the techniques discussed:
🚀 Class reweighting in production at OpenAI
OpenAI was trying to solve the problem of bias in training data of the image generation model DALL-E 2 [1]. DALL-E 2 is trained on a massive dataset of images from the internet, which can contain biases. For example, the dataset may contain more images of men than women or more images of people from certain racial or ethnic groups than others.
These snippets offer insights into how well-known companies grappled with data imbalance and what strategies they adopted to effectively navigate these challenges. For instance, the tip on OpenAI’s approach with DALL-E 2 sheds light on the intricate balance between filtering training data and inadvertently amplifying biases. Such examples underscore the importance of being both strategic and cautious when dealing with imbalanced data. To delve deeper into the specifics and understand the nitty-gritty of these implementations, you are encouraged to follow the company blog or paper links provided. These insights can provide a clearer understanding of how to adapt and apply techniques in varied real-world scenarios effectively.
