Hacker News Reader

Comment by uoaei | Hacker News Reader

Visual sharpness at the expense of wider-scale coherence (see: sliding/floating walking woman in Tokyo demo or tiny people next to giant people in Lagos demo) seems to be a local optimum consistently achieved by today's SOTA models in all domains.

This is neat and all but mostly just a toy. Everything I've seen has me convinced either we are optimizing the wrong loss functions or the architectures we have today are fundamentally limited. This should be understood for what it is and not for what people want it to be.

View on HN

>Visual sharpness at the expense of wider-scale coherence (see: sliding/floating walking woman in Tokyo demo or tiny people next to giant people in Lagos demo)

Wider-Scale coherence is still much better than previous models and has consistently been improving. It's not "visual sharpness at the expense of coherence". At worst, the models are learning wider-scale coherence slower.

Not everything is equally difficult to learn so it follows that some aspects will lag behind others. If coherence weren't improving you might have a point but it is so...

Scaling laws operate in the limit but eventually practical considerations dominate. There's a lot we haven't yet fully appreciated about biological vision and cognition -- and indeed, common sense as regards sensible video generation and processing -- that have not made their way into this kind of model. NeRFs are interesting and I hope to see more from that side of things in the coming months and years.

Nature is great and all but looking to it as an example of a lack of scaling and brute force is a bit ridiculous.

Your vision is hundreds of millions of years in the making.

Yes and in that time we've learned some important lessons that it would be unwise to ignore, e.g. comprehension of 3D geometry despite 2D input visual data.

Did you just recreate the infamous DropBox comment? https://news.ycombinator.com/item?id=9224

That seems like quite the reach, but we will see if it really is just "all you need is scale".

Seems like every big jump in improvement someone says "its just a toy"

Its like we keep moving the bar

If I put goalposts ahead of the ball every time I kick it, can I be said to be accurate? If we don't specify what the actual goal is before calling something an improvement, can we say we are doing anything meaningful?

Reply on news.ycombinator.com