Hacker News Reader

Comment by argonaut | Hacker News Reader

A nuance that is underappreciated by a lot of people / the media is the degree to which any of the AI generates X models (faces, drawings, fills in photos, increases sharpness, etc.) simply copy and paste and interpolate from the training set.

Unlike general supervised learning problems, for many of the deep learning "generative" models that get posted to HN regularly, there is no objective "test set" to measure generalization, so it's extremely easy to claim the model has learned something. When we see a cool demo / audio sample / pictures, how do we know the model hasn't just simply interpolated a bunch of training examples? In many cases this is quite clear, like when you start seeing cats everywhere in the generated imagery. It's very hard to ferret out the BS with these models.

View on HN

Actually, there is a formal method to evaluate generative models. All you need is a validation set. Then you need to find what the odds are that the model would have generated your validation set. That probability is the one you want to maximize [0]. In order to be able to calculate or estimate this probability, you need to make some amendments to your model architecture. But for instance PixelCNN's know the probability exactly, variational auto-encoders have a lower bound on the probability, etc. It is mainly the generative adverserial networks (GANs) which cannot. And this is the reason research in the latter is limited, despite the remarkable visualizations it produces.

But the judging of models by their visualizations is something which is still done too often, and it annoys quite some researchers in the area that those papers still pass review at machine learning conferences. They should be sent to computer graphics conferences, because that is what they actually do /vent.

[0] https://arxiv.org/abs/1511.01844

Yes, I'm mainly talking about generative adversarial models, which have proliferated greatly.

One way to check for overfitting is to compare samples to their nearest neighbors in the training set. Of course that doesn't cover interpolations between two points in the training set, but I would argue plausible interpolations are already something "learned".

A different problem I see with faces in particular though is that our visual system is actually wired to do some really heavy denoising/pattern matching on faces (for example people seeing the face of Jesus on slices of toast), so the generation of faces doesn't actually need to be that good to produce results that seem appealing to humans.

Eh, I'm sure many people copy and paste from Stack Overflow. (Not me ofc, I would never:) )

So I guess we are all out of jobs in a couple years...

we'll still have job but we'll be explaining an AI what we're looking for like it's 5.

that sounds pretty easy to debug though

but the verbosity.... it will be a lot of writing, even for a simple function

> that sounds pretty easy to debug though

Sounds like a nightmare to me. How do you describe a bug to an AI? "Dear AI, that SQL query you did break on my server. Could you improve it and regenerate all your code? ... Nope, still doesn't work. Could you try again?"

and the guy you eventually pay to fix the bug comes back from the bahamas and spends 2 mins to fix it then goes back home with a fat check.

Interesting how you consider interpolating from training data not learning. Isn't that like exactly what learning is? If you get unintended repeating patterns then there isn't enough diverse training data

If there is not enough diverse training data then perhaps you should not call it learning in first place. The whole notion of learning is about algorithm + correct train data meant to make sense of certain kind of test data. All three components matter a lot and it is something that even the experts sometimes seem to miss.

Running a crude algorithm some small set of training data specifically meant to amaze you on specific algorithm have very little information content.

I'm not an expert in deep learning (yet) so someone please correct me if I'm wrong.

Technically I see no reason you couldn't say that meets the definition of learning.

But perhaps that's asking the wrong question. Dictionaries aside, one might decide that a useful standard of "learning" in certain contexts is to have a level of understanding of the subject. And then we ask, does the system have understanding? What is understanding?

A system that only interpolates training data can't logically become consistently better at the task than the data it learns from. Whereas a a sufficiently deep learning system (including, but not limited to, a human) can defeat its master.

Reply on news.ycombinator.com