4 ms·

If I need to show 1 million cat images to train a neural net to see cats, I wonder how a human brain can figure out cats of any kind just by seeing one or two.

If I need to show 1 million cat images to train a neural net to see cats, I wonder how a human brain can figure out cats of any kind just by seeing one or two. Is there something fundamental we are missing in going about building these deep learning stuff ?

bomdo 10y ago
Great question. Even a human brain has to train for ~4-5 months to become interested in shapes (https://en.wikipedia.org/wiki/Infant_visual_development https://en.wikipedia.org/wiki/Infant_visual_development). Once the human brain has been trained for these basic shapes for a while, it is able to quickly break down a new class (i.e. a cat) and recognize similar patterns in other images. This is something that is very similar to the way that training a deep NN works. Also don't forget that the current NN are being trained mainly for photos, not moving images. A brain may recognize a cat by its tail-wagging or fur movements, which is a dimension that is completely missing from still images. Also check out this similar post and discussion: https://news.ycombinator.com/item?id=9247851 https://news.ycombinator.com/item?id=9247851
- vidarh 10y ago
  > Even a human brain has to train for ~4-5 months to become interested in shapes Everyone who wonder how (on a superficial level) grown up humans are so good at learning new categories really should spend time around babies and toddlers and children for this reason... You quickly realise how much training and brain development it actually takes before we're capable of doing much.
  - bpizzi 10y ago
    By far the greatest part of the time I spent with my two toddlers (5 and 3) is spent repeating the same things, again and again, and trying to find shorcuts for learning (which shorcuts might just not exist, afaik). Then one day they start to get it (like "fire burns"), but it's still not there for sure until they experiment it deeply multiple times. The dev in me can't help but see this two little humans as big mighty neural networks who spend their full uptime constantly ingesting tremendous amount of data and being restlessly tuned back by adults and experience :)
    - booli 10y ago
      And this tuning is of utmost important. People often forget that the tuning is human intervention with specific 'specs'. The little baby machine get's quickly tuned to these, good or bad.
maciejgryka 10y ago
Only some billions of years of evolution and a lifetime of knowledge abut how our world looks like! When you see a picture of a cat your brain has a whole lot of context to put the picture into: it looks like some live creature, has eyes, legs, you can tell where it's head is, you can imagine how the fur feels like etc. An untrained neural net has to learn everything from scratch (ha!), from pixel values to "knowing" what a cat is. There is some work on decreasing the amount of data needed to learn, but it's a very tricky subject.
lottin 10y ago
And how many different pictures of cats are there? If you use so many pictures to train your network, the chances are that almost every picture out there will already be in the so-called training set - not the same exact picture, but a very similar one. Then the whole "AI" is reduced to doing some fuzzy matching on a massive database of pictures.
- Sean1708 10y ago
  > And how many different pictures of cats are there? I would guess at a shite-load more than 1 million.
oh_sigh 10y ago
"one-shot learning aims to learn information about object categories from one, or only a few, training images": https://en.wikipedia.org/wiki/One-shot_learning https://en.wikipedia.org/wiki/One-shot_learning
- antognini 10y ago
  There's also the more challenging problem of zero-shot learning, e.g.: http://jmlr.org/proceedings/papers/v37/romera-paredes15.pdf http://jmlr.org/proceedings/papers/v37/romera-paredes15.pdf "Zero-shot learning consists in learning how to recognise new concepts by just having a description of them."
PeterisP 10y ago
You don't really need a 1 million images of specifically cats - there are ML techniques that can look at a million meaningful pictures of different kinds to learn general visual features and then use a much smaller number of cat pictures (possibly just one) as a reference to what do you want to identify. In that sense it's quite similar to a vision system of humans or other animals, which needs lots and lots and lots of early-age exposure to "learn how to see" (which is the hard part); and only after that it becomes possible to figure out cats of any kind just by seeing one or two.
ansible 10y ago
This difference is due to deep learning not being able to build models. Check out the recent paper "Building Machines That Learn and Think Like People": https://arxiv.org/abs/1604.00289 https://arxiv.org/abs/1604.00289 Humans can see one or a few examples of a novel object, such as a cat, and create a fully 3D mental model of it. So we know what it will look like in different orientations and lighting conditions.
- thallukrish 10y ago
  Thanks. Will check out this paper. While in general I felt something similar that it has to be not just reams of data but also some sort of model (meta data) that when combined can produce innumerable combinations more easily.
- TheOtherHobbes 10y ago
  I don't think we have a fully 3D mental model. I think we have a set of core invariants that are constant across a wide variety of conditions. I.e. we can recognise static 2D photos of cats, 2D movies of cats, 3D movies of cats, and real cats in the real world. The invariants are probably relatively simple - a set of head geometries and head feature shapes/distribution, with some secondary colour and texture confirmation. What's interesting is that we can recognise modifiers to the invariants - e.g. a shaved cat is still recognisably a cat, but parsed as "cat without fur." We can also recognise invariants when they're pared down to essentials in cartoons and sketches. https://www.youtube.com/watch?v=R9qdyXCVNVk https://www.youtube.com/watch?v=R9qdyXCVNVk A lot of learning is really just data compression - finding a minimal set of low-resource invariant specifics from a wide range of noisy high-resource inputs.
visarga 10y ago
> I wonder how a human brain can figure out cats of any kind just by seeing one or two. A human child receives images on the retina at 20fps, say ... for 12 hours a day, for many years. That is a lot of training, a lot of images received by the brain. What the kid does when we show it something new is to do a kind of fine-tuning of its neural net where previous visual experience is reused in order to quickly learn new types of objects. It's called one shot learning and it can be done in neural nets too.

bomdo 10y ago

Great question. Even a human brain has to train for ~4-5 months to become interested in shapes (https://en.wikipedia.org/wiki/Infant_visual_development https://en.wikipedia.org/wiki/Infant_visual_development). Once the human brain has been trained for these basic shapes for a while, it is able to quickly break down a new class (i.e. a cat) and recognize similar patterns in other images. This is something that is very similar to the way that training a deep NN works. Also don't forget that the current NN are being trained mainly for photos, not moving images. A brain may recognize a cat by its tail-wagging or fur movements, which is a dimension that is completely missing from still images. Also check out this similar post and discussion: https://news.ycombinator.com/item?id=9247851 https://news.ycombinator.com/item?id=9247851

vidarh 10y ago
> Even a human brain has to train for ~4-5 months to become interested in shapes Everyone who wonder how (on a superficial level) grown up humans are so good at learning new categories really should spend time around babies and toddlers and children for this reason... You quickly realise how much training and brain development it actually takes before we're capable of doing much.
- bpizzi 10y ago
  By far the greatest part of the time I spent with my two toddlers (5 and 3) is spent repeating the same things, again and again, and trying to find shorcuts for learning (which shorcuts might just not exist, afaik). Then one day they start to get it (like "fire burns"), but it's still not there for sure until they experiment it deeply multiple times. The dev in me can't help but see this two little humans as big mighty neural networks who spend their full uptime constantly ingesting tremendous amount of data and being restlessly tuned back by adults and experience :)
  - booli 10y ago
    And this tuning is of utmost important. People often forget that the tuning is human intervention with specific 'specs'. The little baby machine get's quickly tuned to these, good or bad.

maciejgryka 10y ago

Only some billions of years of evolution and a lifetime of knowledge abut how our world looks like! When you see a picture of a cat your brain has a whole lot of context to put the picture into: it looks like some live creature, has eyes, legs, you can tell where it's head is, you can imagine how the fur feels like etc. An untrained neural net has to learn everything from scratch (ha!), from pixel values to "knowing" what a cat is. There is some work on decreasing the amount of data needed to learn, but it's a very tricky subject.

lottin 10y ago

And how many different pictures of cats are there? If you use so many pictures to train your network, the chances are that almost every picture out there will already be in the so-called training set - not the same exact picture, but a very similar one. Then the whole "AI" is reduced to doing some fuzzy matching on a massive database of pictures.

Sean1708 10y ago
> And how many different pictures of cats are there? I would guess at a shite-load more than 1 million.

oh_sigh 10y ago

"one-shot learning aims to learn information about object categories from one, or only a few, training images": https://en.wikipedia.org/wiki/One-shot_learning https://en.wikipedia.org/wiki/One-shot_learning

antognini 10y ago
There's also the more challenging problem of zero-shot learning, e.g.: http://jmlr.org/proceedings/papers/v37/romera-paredes15.pdf http://jmlr.org/proceedings/papers/v37/romera-paredes15.pdf "Zero-shot learning consists in learning how to recognise new concepts by just having a description of them."

PeterisP 10y ago

You don't really need a 1 million images of specifically cats - there are ML techniques that can look at a million meaningful pictures of different kinds to learn general visual features and then use a much smaller number of cat pictures (possibly just one) as a reference to what do you want to identify. In that sense it's quite similar to a vision system of humans or other animals, which needs lots and lots and lots of early-age exposure to "learn how to see" (which is the hard part); and only after that it becomes possible to figure out cats of any kind just by seeing one or two.

ansible 10y ago

This difference is due to deep learning not being able to build models. Check out the recent paper "Building Machines That Learn and Think Like People": https://arxiv.org/abs/1604.00289 https://arxiv.org/abs/1604.00289 Humans can see one or a few examples of a novel object, such as a cat, and create a fully 3D mental model of it. So we know what it will look like in different orientations and lighting conditions.

thallukrish 10y ago
Thanks. Will check out this paper. While in general I felt something similar that it has to be not just reams of data but also some sort of model (meta data) that when combined can produce innumerable combinations more easily.
TheOtherHobbes 10y ago
I don't think we have a fully 3D mental model. I think we have a set of core invariants that are constant across a wide variety of conditions. I.e. we can recognise static 2D photos of cats, 2D movies of cats, 3D movies of cats, and real cats in the real world. The invariants are probably relatively simple - a set of head geometries and head feature shapes/distribution, with some secondary colour and texture confirmation. What's interesting is that we can recognise modifiers to the invariants - e.g. a shaved cat is still recognisably a cat, but parsed as "cat without fur." We can also recognise invariants when they're pared down to essentials in cartoons and sketches. https://www.youtube.com/watch?v=R9qdyXCVNVk https://www.youtube.com/watch?v=R9qdyXCVNVk A lot of learning is really just data compression - finding a minimal set of low-resource invariant specifics from a wide range of noisy high-resource inputs.

visarga 10y ago

> I wonder how a human brain can figure out cats of any kind just by seeing one or two. A human child receives images on the retina at 20fps, say ... for 12 hours a day, for many years. That is a lot of training, a lot of images received by the brain. What the kid does when we show it something new is to do a kind of fine-tuning of its neural net where previous visual experience is reused in order to quickly learn new types of objects. It's called one shot learning and it can be done in neural nets too.