I took Hinton's course on Coursera but I am not yet confident I know the best methods for a given task. I am looking to start with a small project to get more comfortable with machine learning.
If I had say a thousand or so images and I wanted to cluster them, what are the possible methods available to me?
I was thinking of using something like a convolutional neural network (with two streams, one for black and white versions of the image and one for colour) with max pooling layers in-between. Essentially, similar to the one Hinton's group used for ImageNet. Or anything simpler because my purpose is less specific. But I am not sure I know how to use that set up in context of unsupervised clustering. For all I know, this is a dumb way to cluster images and there are probably better methods.
But having no useful experience in the field, I am looking for advice and thoughts on what's the best way to go about this.
Essentially, the goal is to feed it images and have it return a sort of similarity map (like a t-SNE representation, I guess?). So hopefully if I give it 1000 pictures of people and 1000 pictures of flowers, it will cluster them into two massive groups (and whatever sub-clusters it deems necessary).
What are the best modern methods to achieve this goal and what should I look out for?
[link][comment]