3 ms·

> Except a very large fraction of people don't think this way (eg. those with aphantasia) I don't think it's accurate to say that aphants don't have this type

by v64 7y ago

> Except a very large fraction of people don't think this way (eg. those with aphantasia) I don't think it's accurate to say that aphants don't have this type of sensory information at their disposal. I have aphantasia, but I still experience the world through my senses. I may not be able to visualize a cat in my mind's eye, but based on my prior experiences with cats, I know a cat when I see one. If I hear purring, I recognize that as a sound that cats I've encountered in the past have made, etc.

Veedrac 7y ago
Yes, you are capable of pattern matching. The point is that this is a distinct thing from intellectual ability, since we know having orders of magnitude less sensory input doesn't much seem to limit the ability to do high-level reasoning, and know some people don't use it as part of high-level reasoning. This kind of pattern matching is also fairly evidently not all that difficult, since much simpler brains than ours can manage it, as can ML models with caveats (albeit caveats often misunderstood and exaggerated).
- eli_gottlieb 7y ago
  >This kind of pattern matching is also fairly evidently not all that difficult, since much simpler brains than ours can manage it, as can ML models with caveats (albeit caveats often misunderstood and exaggerated). Do tell me, since I'm writing a paper on a related topic, which current ML models can "pattern match" to recognize or generate multimodal (ie: visual, auditory, and tactile) percepts of cats, in arbitrary poses, in any context where cats are usually/realistically found? Or did you just mean that the "cat" subset of Imagenet is as "solved" as the rest of Imagenet?
  - Veedrac 7y ago
    Please try to argue in good faith. I've already said ML models have caveats, obviously I don't think they're perfect or par-human.
    - eli_gottlieb 7y ago
      I think that "perfect" or "par-human" would be a judgement about performance on a set computational task. My caveat is that ML models are usually performing a vastly simplified task compared to what the brain does. But it looked like you were saying they perform "pattern matching" with the same task setting and cost function as the brain, and just need to perform better at it. What's your view?
      - Veedrac 7y ago
        “Not all that difficult” is in the context of the brain, where things tend to vary between ‘pretty difficult’ and ‘seemingly impossible’. I say ML shows pattern matching of this sort isn't all that difficult because progress has been significant over very short stretches of time, without any particular need to solve hard problems, and with a general approach that looks like it will extend into the future. We have this famous image showing progress over the last 5 years. https://pbs.twimg.com/media/Dw6ZIOlX4AMKL9J?format=jpg&name=large https://pbs.twimg.com/media/Dw6ZIOlX4AMKL9J?format=jpg&name=... The latest generator in this list has very powerful latent spaces, including approximately accurate 3D rotations. https://youtu.be/kSLJriaOumA?t=333 https://youtu.be/kSLJriaOumA?t=333 We have similarly impressive image segmentation and pose estimation results. https://paperswithcode.com/paper/deep-high-resolution-representation-learning-2 https://paperswithcode.com/paper/deep-high-resolution-repres... Because you mentioned it, note that models that utilize multimodal perception is possible. The following uses audio with video. https://ai.googleblog.com/2018/04/looking-to-listen-audio-visual-speech.html https://ai.googleblog.com/2018/04/looking-to-listen-audio-vi... For sure, these are not showing off the full breadth of versatility that humans have. I can still reliably distinguish StyleGAN faces from real faces, and segmentation still has issues. These all have fairly prominent failure cases, can't refine their estimates with further analysis like humans can, and humans still learn much, much faster than these models. However, note that (for example) StyleGAN has 26 million parameters, and with my standard approximate comparison of 1 bit:1 synapse, that puts it probably somewhere around the size of a honey bee brain. Given such a model is already capturing sophisticated models fairly reliably using sophisticated variants of old techniques without need of a complete rethink, and the same cannot be said for (eg.) high-level reasoning, where older strategies (eg. frames) are pretty much completely discredited, “not all that difficult” seems like a pretty defensible stance.

by v64 7y ago

Veedrac 7y ago

Yes, you are capable of pattern matching. The point is that this is a distinct thing from intellectual ability, since we know having orders of magnitude less sensory input doesn't much seem to limit the ability to do high-level reasoning, and know some people don't use it as part of high-level reasoning. This kind of pattern matching is also fairly evidently not all that difficult, since much simpler brains than ours can manage it, as can ML models with caveats (albeit caveats often misunderstood and exaggerated).

eli_gottlieb 7y ago
>This kind of pattern matching is also fairly evidently not all that difficult, since much simpler brains than ours can manage it, as can ML models with caveats (albeit caveats often misunderstood and exaggerated). Do tell me, since I'm writing a paper on a related topic, which current ML models can "pattern match" to recognize or generate multimodal (ie: visual, auditory, and tactile) percepts of cats, in arbitrary poses, in any context where cats are usually/realistically found? Or did you just mean that the "cat" subset of Imagenet is as "solved" as the rest of Imagenet?
- Veedrac 7y ago
  Please try to argue in good faith. I've already said ML models have caveats, obviously I don't think they're perfect or par-human.
  - eli_gottlieb 7y ago
    I think that "perfect" or "par-human" would be a judgement about performance on a set computational task. My caveat is that ML models are usually performing a vastly simplified task compared to what the brain does. But it looked like you were saying they perform "pattern matching" with the same task setting and cost function as the brain, and just need to perform better at it. What's your view?
    - Veedrac 7y ago
      “Not all that difficult” is in the context of the brain, where things tend to vary between ‘pretty difficult’ and ‘seemingly impossible’. I say ML shows pattern matching of this sort isn't all that difficult because progress has been significant over very short stretches of time, without any particular need to solve hard problems, and with a general approach that looks like it will extend into the future. We have this famous image showing progress over the last 5 years. https://pbs.twimg.com/media/Dw6ZIOlX4AMKL9J?format=jpg&name=large https://pbs.twimg.com/media/Dw6ZIOlX4AMKL9J?format=jpg&name=... The latest generator in this list has very powerful latent spaces, including approximately accurate 3D rotations. https://youtu.be/kSLJriaOumA?t=333 https://youtu.be/kSLJriaOumA?t=333 We have similarly impressive image segmentation and pose estimation results. https://paperswithcode.com/paper/deep-high-resolution-representation-learning-2 https://paperswithcode.com/paper/deep-high-resolution-repres... Because you mentioned it, note that models that utilize multimodal perception is possible. The following uses audio with video. https://ai.googleblog.com/2018/04/looking-to-listen-audio-visual-speech.html https://ai.googleblog.com/2018/04/looking-to-listen-audio-vi... For sure, these are not showing off the full breadth of versatility that humans have. I can still reliably distinguish StyleGAN faces from real faces, and segmentation still has issues. These all have fairly prominent failure cases, can't refine their estimates with further analysis like humans can, and humans still learn much, much faster than these models. However, note that (for example) StyleGAN has 26 million parameters, and with my standard approximate comparison of 1 bit:1 synapse, that puts it probably somewhere around the size of a honey bee brain. Given such a model is already capturing sophisticated models fairly reliably using sophisticated variants of old techniques without need of a complete rethink, and the same cannot be said for (eg.) high-level reasoning, where older strategies (eg. frames) are pretty much completely discredited, “not all that difficult” seems like a pretty defensible stance.