Profile cover photo
Profile photo
Ferenc Huszár
207 followers
207 followers
About
Posts

Post has attachment
I wonder if people here have any thoughts on this:
In my latest post I talk about Gaussianization of autoregressive generative models of images. I think this is an interesting technique in itself that's perhaps not very widely known.

Main advantage though is that one can use this to derive closed-form objective functions based on adversarial training or kernel moment matching. Both of these methods currently rely on sampling, which limits their applicability to relatively low dimensional data.

Warning: the post is a bit verbose, a little bit write-only, but hopefully you'll find the ideas useful.

Post has attachment
My latest blog post, this time explaining Gaussianisation in autoregressive image models, with some new ideas at the end.

Key idea: One can combine Gaussianisation with adversarial/moment matching objective functions, resulting in a closed form objective that does not require sampling. Revolutionary! :)

Warning: this is long, mostly write-only dump of thoughts, comments are welcome - as always (my disqus plugin is being an a...dversary)
Add a comment...

Post has attachment

Post has attachment
Shall we change the comminity logo to something more interesting, +Azeem Azhar . Something like this?
Photo

Post has attachment
I'll need to have a closer look at David Barber's book [available free online] I looked at early drafts and it looks like it's put together quite well.

Post has attachment
Add a comment...

Post has attachment
Bregman divergence: Peter and Quentin live on a convex hilltop. How high does Peter have to jump to see Quentin?
Photo
Add a comment...

Post has attachment
I don't have my homework because <%= reason %>.
Add a comment...

Post has attachment

Post has attachment
If high quality lectures and course material become available online, what's going to be the new role traditional universities can play in this ecosystem? Which universities are going to adopt to the new situation? I think universities where the focus is on small-group experience, personal curriculum development, expert supervision, and overall experience like +Cambridge University or +Oxford University may survive (again) and conserve themselves as leaders if they recognise how their strengths complement new technology. Thoughts please.
Add a comment...
Wait while more posts are being loaded