How did a bot come up with a neural network that interprets the popular meme images on Twitter?



460,000 '

neural net guesses memes (NNGM)' is a bot that interprets popular images that many people like and RT on Twitter with a neural network and guesses 'what is in it' and tweets it. It is popular as a follower. Tech reporter Hayden Field talks to the author about why such a bot was created.

The creator of the viral 'neural net guesses memes' Twitter account explains how it works
https://www.morningbrew.com/emerging-tech/stories/2022/02/16/the-creator-of-the-viral-neural-net-guesses-memes-twitter-account-explains-how-it- works



GitHub --riscmkv / NNTwitterBot
https://github.com/riscmkv/NNTwitterBot

An example of NNGM's tweet looks like this. The following is a story that Mr. Neumann also likes, and it is a picture of a large amount of eggs in the refrigerator, but it is estimated by the interpretation of the neural network that it is 'ping-pong ball with 99.99% certainty'.



The author is Glen Neumann, a computer engineer. When I was in college, I encountered machine learning, and through the task of interpreting animal images to judge whether it was a dog or a cat, I became interested in the part such as 'how the neural network thought and came to that conclusion'. ..

After graduating from university, Neumann got a job unrelated to neural networks, but thought that 'it would be a good opportunity to play with neural networks' and 'it would be a good opportunity to learn about Twitter API', so he started developing Twitter bot NNGM.

The code itself was completed in a few hours using PyTorch 's trained ResNeXt model, and the rest was struggling with the Twitter API part. One of the reasons I chose ResNeXt as the image recognition model was that I was the winner of the 2016 competition to create an accurate neural network using the dataset ' ImageNet ', but PyTorch of Python library was preliminarily used for ResNeXt. He said he had a trained model and didn't have to code and train it himself.

NNGM first buzzed in October 2021 with the following tweet, which estimated that the photo 'A toy tank has a stuffed animal of Reimu Hakurei' was 'snowmobile with 37.05% certainty'. It was said that this tweet arrived at the community where fans of the Touhou Project gathered and became a hot topic.



While the account was small, it seems that they processed the 'theme' photos sent by DM one by one, but eventually they started to process by scripting, and the number of DMs increased too much. The processing by the script was too late, and a post-only site was created.

Mr. Neumann, who wanted to show the greatness of the neural network, was convinced that the guess was wrong, 'It can not be helped as long as the meme image is loaded,' rather, 'Why did the bot think so?' I'm a little scared when I find out that I made such a guess. '

In addition, although it is a bot that sometimes misses a big guess, there are people who love the degree of deviation and get excited when the guess hits, 'When I see people who are excited about guessing about meaningless memes, It feels really fun, 'says Neumann.

In addition, Mr. Neumann's favorite tweet selection is the following lineup.

The cheese-filled hamburger is 'pineapple with 99.3% certainty.' Mr. Neumann was convinced that 'the condition of this cheese is certainly pineapple-like.'



Is it because the company name 'THE SLIDING DOOR COMPANY' written on the glass pulled the double-swing door to 'sliding door with 36.57% certainty'?



Not only is the guess wrong, but there are times when you can see a photo that you don't know that a pistol is buried in a guacamole , 'guacamole with 99.33% certainty.' I think you can see that this is exciting.



Related Posts:

in Web Service, Posted by logc_nt