Hacker News Reader

Comment by alextgordon | Hacker News Reader

This shows the folly of modern approaches:

> Pornographic images, for instance, are frequently identified not by the presence of particular body parts or structural features, but by the dominance of certain colors in the images

Neural networks are unsuccessful because they are not detecting anything. They're just making statistical correlations that happen to be right for the specific set of training data they have received. There's no robustness whatsoever. Artificial gullibility.

If we want to teach computers to think, then we have to teach them to model the physical world we live in. Plus the emotional, social world.

And even that is entirely separate from heavyweight abstract reasoning. Just because a computer can pass the turing test, doesn't mean it can spontaneously diagnose your medical problems, design a suspension bridge or play Go.

View on HN

You make some pretty strong attacks on the state of the art approach to computer vision. Could you please provide citations?

As someone working on neural networks, although not really on computer vision, I do not believe several of the things you said are true.

> > Pornographic images, for instance, are frequently identified not by the presence of particular body parts or structural features, but by the dominance of certain colors in the images

> Neural networks

This is definitely not what state of the art neural networks for computer vision do, although I'm sure there are many models out there that do function in that manner.

We know that neural networks learn to detect higher level features. For example, see the excellent visualizations in Matthew Zeiler and Rob Fergus' Visualizing and Understanding Convolutional Networks paper. (http://arxiv.org/pdf/1311.2901.pdf)

> Neural networks are unsuccessful because they are not detecting anything. They're just making statistical correlations that happen to be right for the specific set of training data they have received.

Are you saying that the correlations they learn aren't true on non-training data?

If so, that is false. Any competent person in ML tests on different data than they train. Well trained neural networks generalize to new data very well. That's why we use them.

> There's no robustness whatsoever. Artificial gullibility.

Are you referring to the adversarial counterexamples section of Szegedy et al's Intriguing properties of neural networks paper? (http://arxiv.org/abs/1312.6199)

This paper shows that one can craft noise which will fool a neural network. But it isn't noise that is found in real life. It's created by an optimization process which is basically looking for ways to fool the network. Basically, it discovers an optical illusion for the network.

(From the manifold perspective, it moves us in a direction perpendicular to the manifold.)

> This is definitely not what state of the art neural networks for computer vision do, although I'm sure there are many models out there that do function in that manner.

I was just quoting the article really.

Neural networks are moderately good at replicating the part of the brain that identifies visuals, because that is just statistics and neural networks are a slightly awkward way of doing statistics.

But only up to a point, because humans don't sit around all day looking at static images. We humans move around, and the objects we are classifying move. Our world is 4-dimensional, not 2-dimensional.

Perhaps a neural network can tell what a cat looks like, but can it tell the difference between a dead cat and a sleeping one, by watching it breathe?

Neural networks are not good at replicating the parts of the brain that deal with physical properties of the world, because that is not a statistical operation. It's an algorithmic one.

Unfortunately for them, such physical properties are not an afterthought but an inseparable part of the task they are attempting to achieve.

>Perhaps a neural network can tell what a cat looks like, but can it tell the difference between a dead cat and a sleeping one, by watching it breathe?

Yes, yes it can: http://arxiv.org/abs/1411.4389 (Long-term Recurrent Convolutional Networks for Visual Recognition and Description) It's certainly surprising how much of perception can be reduced to classification.

But you can tell porn is porn based on a single 2d image.

Actually, humans often struggle to even agree on what is porn.

Well you can't expect a computer to solve that.

This is what the brain does, too. All inductive reasoning comes down to statistical correlations. The human brain has NEVER had "robustness".

> This is what the brain does

Good to know you know what our brain does, science hasn't figured out yet but you have...

Well, I don't know it, but if my brain operates at a higher certainty it certainly doesn't reach my thoughts.

By robustness I mean the brain is pretty hard to exploit. You can't trick it into thinking an image is pornographic just by changing the color balance.

> The brain is pretty hard to exploit.

No it's not, just ask any advertiser.

> You can't trick it into thinking an image is pornographic just by changing the color balance.

If the only data someone has about an image is its colour balance and you asked them to classify which they think are pornographic, then yes you can.

It's a pretty stupid idea to impose such limitations though, on humans or ANNs. Even if you did only care about colour balance, you don't need an ANN to tell you that; just scale the image down to 1 pixel x 1 pixel, then look at its colour.

> > By robustness I mean the brain is pretty hard to exploit. You can't trick it into thinking an image is pornographic just by changing the color balance. > No it's not, just ask any advertiser.

One of the great things about humans is that they can consider the context of a remark, although apparently this is not universal.

Maybe it's an AI.

There's an entire subreddit devoted to a technique for fooling the brain into thinking an image is pornographic by overlaying patterns that obscures clothed sections [1] (and at least one more about more generic optical illusions)

Not quite the color balance, but how anyone can assert the brain is hard to exploit is beyond me.

[1] it's this technique - links is probably nsfw: https://richardwiseman.wordpress.com/2010/09/07/naked-women-... - I don't remember the name of the subreddit.

The phenomena outlined in http://en.wikipedia.org/wiki/Optical_illusion would seem to instantly belie the claim in your first sentence.

> By robustness I mean the brain is pretty hard to exploit.

As others have pointed out, the complete opposite is true.

http://4.bp.blogspot.com/-dUHs2FI_izg/UTC7OQLDy7I/AAAAAAABhJ...

https://richardwiseman.files.wordpress.com/2010/12/161cq.jpg

http://clampd.com/wp-content/uploads/2014/07/optical-illusio...

I am downvoting you because you are factually incorrect on several points, while maintaining an aggressive, self-important tone. Neural networks are incredibly successful. The state of the art in computer vision performs /as well or better/ than human beings for image classification.

> Just because a computer can pass the turing test, doesn't mean it can spontaneously diagnose your medical problems, design a suspension bridge or play Go.

I think you may have too narrow a view of a turing test. As part of any reasonably tough turing test would be the ability for the test subject to learn something new and apply it.

A computer might be able to learn something new. That doesn't mean it'll be particularly good at it.

Our goal is presumably to design computers that can do the specialised jobs that people do. A computer that can pretend to be a novice is not much use in the grand scheme of things.

Reply on news.ycombinator.com