[aw3c2]

This is a complete joke, right? What is better about those results than a simple "image + headline + random bottom line" algorithm?

[wodenokoto]

Judging from the url posted in an earlier top thread, this might be a student report.

https://web.stanford.edu/class/cs224n/reports/6909159.pdf

[ekianjo]

Exactly. Memes are funny because they make meta references that are culturally relevant or simply attach absurd bottom lines. It's highly unlikely a deep neural network can model anything like that.

[dmschulman]

Considering most deep learning results are interpreted as absurd/bizarre, I don't think the machine will have much difficulty intentionally or unintentionally emulating meme culture.

[vertexFarm]

That was my thought. They need to crank the noise way up and aim for some surreal memes, not these ancient fossilized memes from 2010.

[stochastic_monk]

I think the image needs to be an input somehow. I imagine running an image classifier (e.g., YOLO9000) to extract “pretrained” features and making those values inputs into a modified LSTM could allow learning to synthesize text and perception. I’d suggest learning new image embeddings (training a neural network to extract image features from scratch), but it’d be difficult to get enough images/enough different images.