I think Brooks' Cog initiative was an attempt to 'ground' the robot's perceptions of the physical world into forming a rich scalable representation model. But it looks like that line of investigation ended ~2003 with Brooks' retirement. Too bad, given the seeming suitability of using deep nets to implement it.
I agree. Google has come out with some papers that are, to put it harshly, basic gluing together of DL models followed by loads of training on their compute resources.
How about "They require solutions to problems that the average human mind shies away from attempting to comprehend..." (I assume the rest of the argument is "thus making it hard for anyone with the problem your product solves to realize they even have that problem.")
How so? No one really has any idea how, say, NNs trained on machine vision are working. Hand designed programs to do machine vision don't even get a tenth of the accuracy of NNs. That's just one domain. And one that has been very well studied by humans for decades with little progress.
AI solves problems that are vastly too hard for humans to figure out.
> No one really has any idea how, say, NNs trained on machine vision are working.
False. Plenty of work has been done to show what exactly is going on inside NNs including some very impressive visualizations.
> Hand designed programs to do machine vision don't even get a tenth of the accuracy of NNs.
And that's false too. Hand designed programs to do machine vision can get significant fractions of the accuracy of NNs and sometimes even perform better and/or faster depending on the domain.
It is a lot more work to write that code though and a NN might clue into features that you weren't aware even existed.
Sure there are some pretty visualizations of what images turn on a neuron or whatever. But no one could hand code an algorithm that does the same job that's even half as good.
>And that's false too. Hand designed programs to do machine vision can get significant fractions of the accuracy of NNs and sometimes even perform better and/or faster depending on the domain.
Source? A few years ago people thought it was amazing an algorithm could recognize cats 12% of the time. And even the best non-neural approaches use some form of machine learning.
>a NN might clue into features that you weren't aware even existed.
Your statements are nonsense. Yes, we do have a good idea of what goes on inside a NN trained on images and Yes hand designed programs exist and work and sometimes work very well.
It's just that NNs are rapidly moving ahead and maybe some time in the future both your statements will be true but today they are not.
Recognizing cats is not the sum total of machine vision problems and those 'pretty visualizations' go a lot further than what you give them credit for.
I've been doing nothing but reading up on machine learning and working through piles of example code and models for the last couple of months and I think based on what I've seen so far that we are definitely headed in the direction that you suggest. At the same time we aren't there yet and such hype does not help.
Keep in mind that the domain of machine vision is extremely broad and much wider than Kaggle competitions and that saying 'we have no idea' when clearly we do tries to sketch NNs as magic and they're not.
If anything the present generation is best characterized as an optimization of what you could have done for many years, the only reason they are the way they are is because we'd like to have our results in a reasonable time. That makes applying NNs to large datasets possible and that in turn is what helps to solve some of these problems.
>Yes, we do have a good idea of what goes on inside a NN trained on images
I'd love a source for this. The last time I checked I didn't find much. Sure we know the first few layers do stuff like edge detection. But the higher layers are much harder. And even if you know images are activating them, writing code that can accomplish the same task is much harder.
>It's just that NNs are rapidly moving ahead and maybe some time in the future both your statements will be true but today they are not.
The best NNs have already surpassed human accuracy at imagenet. That's insane! No one expected that result a few years ago. Yes object recognition isn't the only task in machine vision, but it's pretty representative.
I will do some digging for you tomorrow, it's 3:30 am and I really should get some sleep.
> Sure we know the first few layers do stuff like edge detection.
Yes, basic primitives.
> But the higher layers are much harder.
Well, those higher layers are built up out of the primitives the lower layers contain. So you get things that are conceptually easier to understand rather than harder.
> And even if you know images are activating them, writing code that can accomplish the same task is much harder.
That I completely agree with, I've basically gone that route in the last couple of months.
> The best NNs have already surpassed human accuracy at imagenet.
Yep. It's extremely impressive.
It's a watershed moment in the history of computing and half of us are still asleep or in denial. Which is fine by me. But please don't hype. It isn't magic and it isn't perfect - yet.
>> Hand designed programs to do machine vision don't even get a tenth of the accuracy of NNs.
"A tenth of the accuracy" of a 100% accurate model is 10%. So you're basically claiming that nobody had achieved anything with machine vision before NNs (you probably mean AlexNet in particular), which is patently absurd and completely unhistorical.
Why is it that this subject attracts the attention of so many people who know so little of what they're talking about?
In real world datasets a single class may only be in 1% of the data to begin with. The vast majority of images on the internet do not contain cats. So making a cat recognizer that is even 10% accurate is a significant challenge, simply because of the low base rate of cats. Imagenet in particular has 1000 equally balanced classes, so each class makes up 0.1% of the data. And 10% accuracy is way better than chance.
Additionally very few machine vision systems are completely hand designed. They all use some form of machine learning and optimization. Just not necessarily NNs.
Classification is only one problem in machine vision. As in, not even 10% of the whole.
Edit: Btw I just had a look to confirm and what you're saying about "1000 equally balanced classes" is inexact, to put it midly. ImageNet has 100,000 SynSets (as in WordNet) and aims to have about 1000 images illustrating each. Individual images may illustrate several synsets- you don't get one image per class, say.
Also: AlexNet was not a cat recogniser and having unbalanced classes is a problem for machine learning, but not a dealbreaker, per se.
>Classification is only one problem in machine vision. As in, not even 10% of the whole.
Classification is extremely representative of the challenge of machine vision. Good classifiers can often be adapted to do localization and other tasks, e.g. by scanning the image to see what parts contain an object.
The Imagenet Large Scale Visual Recognition Challenge has 1000 classes which you can browse here: http://image-net.org/challenges/LSVRC/2015/browse-synsets I forgot that Imagenet itself has more classes than that. But it's best known for this specific competition of 1000 classes, and everyone uses that as a benchmark.
>freaking AlexNet was not a fraking cat recogniser.
I never said it was... It's actually a dog recognizer mostly : )
I'm curious, did you not hear the feedback? I've been looking at how writing is miscommunicated and misheard and I would be interested in understanding what you heard. I couldn't find an email address in your profile here or on medium; Hopefully you'll see this and respond (in email is fine, mine is in my hn profile, I won't share it with anyone else).
We've developed abstractions to help reason around complexity, high dimensionality, and the problems that can be solved with each. My intention with that particular sentence was more around the actual navigation of that space. However, I didn't put much effort into that paragraph and intend to revise it when I have a bit more time.
Just added my email to my profile - feel free to reach out if I'm missing something or you'd like to discuss more.
Besides graduating from your fast.ai course, what were the other qualifications of those Google Brain applicants? I'm imagining they would have, or be in the process of getting, an MA/PHD in non-AI-related area.
