But detecting memes is complicated: The layered combination of images, text and sometimes ironic context makes them notoriously difficult even for not-safe-for-work (NSFW) AI systems to read automatically. The explicit, with humorous or sarcastic content makes it more difficult to detect memes. In 2023, it was determined that nsfw ai could detect explicit content at a rate of over 90%, but the accuracy dropped to roughly 65% when processing memes because there were text overlays and visual distortions. If they have text, memes often feature large letters and colors combined with irony from simple content detection models that were not designed to handle the more complex rules of meme creation.
Performing Text Analysis within MemesText analysis presents yet another obstacle in memes. Previous ai models use optical character recognition (OCR) to read text from images, however meme texts are typically presented in stylised fonts or with image noise on top. When scanned, this can drop by 25%, especially so if mixed in with non-standard vocabulary or slang. These platforms review millions of memes every day, and their high false-positive results that inadvertently mark innocent memes as nsfw tend to increase the appeals from creators wanting those automated moderation decisions reversed by up to 10%.
Developers can use convolutional neural networks (CNNs) and natural language processing (NLP) layers to ‘look’ at the image component when reading a sentence in the text. The algorithms sift through memes and their parts like captions, pixel groups or even creating canonized format templates for better filtering accuracy. Similarly, in 2022 Twitter reported a 40% improvement of their ai’s ability to recognize potentially explicit memes through using over half a million unique meme samples for model training which covers both clean and dirty formats.
This is one that Elon Musk famously weighed in on by saying “ai humor detection: about as hard as teaching a robot to love” His words reflect the devilish puzzle that is teaching nsfw ai to correctly understand memes. Uncovering memes needs to be more than just detecting the and active content descriptor, it requires an ai with a sense of humour that can also understand sarcasm, tone, and cultural references as well. Meme evolution complicates the ability of ai to sufficiently identify large patterns due to rapid memes format changes which can evolve faster than an average training cycle.
The meme detection capability of nsfw ai is still quite immature even though it has come a long way since the invention. This process was done every 3 to for months in order that the training datasets were recently kept updated with new meme formats. Costing tech firms at least $100,000 per cycle of retraining this nsfw ai that works on meme-specific images calls for some big-money investment. But those costs are the price we pay for ensuring that online spaces remain safe and fun places you can go to connect with fellow humans. For more about how nsfw ai models meme detection and recent progress on the task, you can check out our work at: Monsoon 2021 Update for Meme Detectionavenuesusc / nsfw AI-Blog DocumentsDetection of Memes in Online Imageboards using Deep Visual ExplanationMining Leftist Websites to Frame Conservatives with Home Grown TerrorismStats Settingsdirectories0dotsBlogs + Moreermalinksgithub.comangel-moreno_equal_digitalsexampl…arXiv-Check back next time!