Show HN: Cornea AI – Using Deep Learning to Predict Photo Popularity | Hacker News Reader

Show HN: Cornea AI – Using Deep Learning to Predict Photo Popularity

cornea.ai

45 pointsparths15 comments

This is very interesting. I cannot stop thinking that this is the logical next step of social media, or if you want to make the circle a bit bigger - photography; or even bigger - self representation.

People strive to be perceived in a certain light and be approved for this egocentric perception of oneself. First it started with favorable pictures that you deemed fitting. Then along came filters, which "enhance" certain features of your pictures. Now Cornea AI - which guesses the approval rate of your pictures.

I somehow find it beautiful. But then again I feel a bit sad that humans have come thus far...

You summarised our vision very well, we want to democratise creativity so you never have to guess what will work on social media

I'm sorry but what does this have to do with democratization?

Could you plug this into one of those image generation algorithms and generate a potentially popular image from scratch?

Side note: I can imagine high speed ad networks in the future using this to deliver custom images in real time to users based on their profiles. Tweaked of course based on the interests of their extended network and the likelihood of going viral within that network.

Hi, That can be done, but image generation algorithms require a bit more time to be good enough for such a thing. Two major bottlenecks (I am assuming you are talking about GANs) are: 1. The algorithms are still not too good for hi res images. 2. The algorithms work very well on images where fine grained features are not that important. (For example images of nature), but they introduce unwanted features otherwise (Eyes looking other ways etc.). We might be all ready by next ICLR/NIPS though, cannot say.

Very interesting. Can you please talk about the algorithm you use to predict the popularity score .

A neural network has been trained to differentiate between popular and non-popular images and give like scores. This network was optimised to understand the features that made a photo popular on social media. Part of the work was inspired by Karpathy's blog on selfie (http://karpathy.github.io/2015/10/25/selfie/). However, our algorithm works on different type of images and not just selfies.

To add a bit to Parth's answer it has been trained using metric learning type losses and normal softmax combined. There is a next version of combined low/high level features combined algorithms we are slowly going to update into.

Maybe someone should do an app that writes facebook posts for you, predicting what will be the most popular.

Or Reddit -- since the populaity of all Reddit posts is known it would be a good training set for such an AI :-)

Absolutely! With visual content being predicted for popularity, a model can be made which produces quality content based on the training data of existing popular content from different sources.

Interesting technology, not much of a photo person but would like to know how are you building the dataset considering there are so many different types of images that can trend at any point of time?

Score seems to work on some images but did not work on my dog's photo - how do you guys evaluate the accuracy?

We have trained our model on public photos which were popular/trending at some point of time. We fed these images to a deep CNN (Convolutional Neural Network) which started to recognise features that made photos popular. What we realise in the process that these features do change with time so we have added a temporal component to our training set to ensure our model is relevant. It is currently optimised more for human photos and travel images so maybe in our next iteration, we can predict pets photos as well.

Do you consider a simple web interface where I can upload a photo?

As of now, the focus is only on the mobile apps. Though,in future we are planning to release an API with a web demo for the same.

Reply on news.ycombinator.com