My First Kaggle ConvNet: Getting to 3rd Percentile in 3 Months

Intro:

'The Diabetic Retinopathy challenge on Kaggle has just finished. The goal of the competition was to predict the presence and severity of the disease Diabetic Retinopathy from photographs of eyes. I finished in 20th place using a Convolutional Neural Network (ConvNet). In this post I’ll explain my learning process and progress as I implemented my first ConvNet over the last 3 months. Throughout, I’ll link to the implementations in my code, which is available on github for anyone who wishes to replicate my score.'

So what won? Did they use conv nets?

Still active: https://www.kaggle.com/c/diabetic-retinopathy-detection

Looks like a fun problem. I've never had much luck with NN for anything other that feature extraction.

It just finished, I got 4th. I used convnets. With the correct initialization and data preprocessing they actually are quite easy to train.

Neat. Congrats! I'm going to have to learn me some ConvNets.

From the winner's publications list: http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic-...

Yes, they must've.

Nothing new. Deep learning is hyped up to 90% these days.