The NSFW AI works by learning from data through supervised learning, which means it is trained on large datasets that have labeled examples of both SFW and OS images. This usually involves millions of human-annotated images and videos. Such as an AI model trained on data set size of more than 10 million images where deviant sexual activity is involved in only about 20% out of them. These labeled examples train the AI to recognize patterns, features and contexts that differentiate explicit content from non-explicit material.
This is then used as the training input for neural networks, particularly a type of convolutional neural network (CNN) which performs well on image recognition tasks. CNNs review images at the pixel, color, shape and texture level in layers to learn how progressively higher filters detect objects/features that might be associated with NSFW content. The AI becomes better at weeding out the content while processing more data, completing training cycles that can take weeks (depending on dataset and computational resources). Which can become faster with high end GPUs — those have training speeds of over 50K images/s.
Use cases in the industry reveal some of the possibilities as to how this technology could be used. Every month, there are billions of pieces of content on sites like Reddit and Tumblr being moderated by NSFW AI. While the technology is incredibly powerful, it still has its limitations. Similar instances happened in 2019 (even when a photo was just of breastfeeding!), and controversies about the bias with which algorithms operate started to bubble up. Here are two examples that show why fresh data needs to feed continuous updates and re-training.
This is where companies implement continuous learning and at Google the learning cycle really comes into play. This means constantly feeding the AI fresh data so that it can adjust to new trends (ex. changing visual styles or formats). Studies in effectiveness reveal that retraining the models with fresh data every quarter advances detection figures by up to 15% higher and mitigates false positives as well as negatives.
Data is, quoting AI researcher Andrew Ng — “the new electricity”. It is a saying that goes the more heterogeneous and pertinent your data; better are your results! It demonstrates the need for good datasets when using NSFW AI as a training model. However, but high data quality is not just the god use of quantity in terms here it referred to ensure that our training set doesn't include only edge cases nor ambiguous content.
Also because of the nature mentioned, nsfw ai use deep learning algos which keep getting updated as and when new data keeps flowing in an ethical way to help balance accuracy vs ethics. The learning loop is ongoing and constant data input, model tuning, validation on content grounded in the real world are vital to ensure that moderation happens effectively across platforms.
