Hacker News Reader

Comment by nightski | Hacker News Reader

What is the difference between a data set and a raw sensory data stream? More specifically, isn't a raw sensory data stream just a data set? I think you are getting hung up on semantics.

Or is it just the time correlation that interests you? Because some of these data sets are very likely to indeed be time correlated. Like a video/audio data set for example.

View on HN

Yes in the most general sense, a "data set" could be anything, but I'm talking about the highly curated and labeled data sets that are used to train contemporary ML systems.

You can feed a million images that are labeled "cat" or "no cat" to one of these systems and it can achieve a human-level proficiency at identifying cats in images. But, it won't be able to do anything other than identify cats, it's far too narrow to be considered intelligent in any way.

If you can feed a series of timestamped photos to a system, basically an unlabeled arbitrary video stream, and you can demonstrate that it formed some notion of what a cat is, that would be very interesting indeed.

Have you ever seen a child learn how to speak? He also needs loads of "labelled" data to initially learn about concepts.

The data is not labelled in the same way. The child has still acquired their knowledge solely from sensory experience. How do they know to apply the spoken label "mom" to the recurring pattern in their visual data stream?

With modern data sets and labelling, so much of the problem domain is deeply hard-coded into the system. It doesn't have to learn what letters and words mean and how to identify them in a totally arbitrary* visual or audio stream. It just gets a relatively minute amount of structured data that it has an embedded understanding of what to do with.

Sure there are things like OCR and speech-to-text, but I don't think you could just run your streams through those, there's just so much subtle information loss. In order to make AI that really has a chance of reaching what humans would call intelligence, I firmly believe it has to make meaning out of some kind of raw sensory experience analogous to ours.

*Ok, not totally arbitrary, a human would not learn language from a video of a forest, and that's where the "labelling" comes in, from observing other people using language, but a child can still learn any language from just that, and the labels are often vague, inaccurate, contradictory, complex, abstract, etc. There's no master training set with the right answers, you have to decide for yourself. And humans were also capable of bootstrapping language from nothing. I just don't see modern supervised learning systems ever doing things like that.

Actually you can learn a lot of stuff unsupervised, and then just use supervised learning to extract the information from within the net. So you can just throw huge amounts of letters at a deep net then later look at the activation at the top and say "Ok, so when neurons X, Y and Z are on it's when it's looking at a 1". I'm pretty sure you can just connect two at the top and have it say the number it sees.

> In order to make AI that really has a chance of reaching what humans would call intelligence, I firmly believe it has to make meaning out of some kind of raw sensory experience analogous to ours.

And why must it be able to do this from scratch? Why must it be hampered with the same limitations we have?

Unlabelled datasets are valuable in themselves.

Self-Taught Learning: Transfer Learning from Unlabeled Data

http://www.andrewng.org/portfolio/self-taught-learning-trans...

I would disagree on this point, humans unlike current AI systems can learn from one or two data points, especially at easier tasks like identifying cats. Current AI algorithms need huge labeled data sets for solving narrow problems so one needs to build more generalization ability to our current AI systems.

That's because we are already VERY pretrained on very similar tasks. You should look into transfer learning and one shot learning for examples of techniques that in practice do something very similar.

Yes, I have heard of transfer learning and used them in practice. Very powerful but still primitive. One-shot learning techniques are still to mature but I agree with you that technologies like these will reduce our reliance on datasets and make the AI algorithms learn in more human like way.

Are you comparing a child learning to speak with a set of conditional statements?

I'm personally most interested in approaches to AI not based as heavily on dataset-collecting. There's a now almost standard method of: 1) first curate a nice, large dataset (often explicitly labeled by humans), then 2) carefully engineer a model architecture that has good performance on the problem represented by that dataset. Example: dataset is ImageNet, problem is tagging images. Obviously there's a lot of practical value in that, and it's a perfectly reasonable thing to research and use, but that's only one specific kind of inference.

this all seems quite naive of the literature in the field. what other "kind of inference" do you wish he would pursue?

Quite possible. I do have a recent PhD in AI, but it's a big field and I can't claim to know more than a small percentage of it.

The problem with the dataset-first approach is that humans are still providing a lage part of the intelligence: defining the domain, defining what good performance on it looks like, carefully designing model architectures, collecting and labeling large datasets, etc. This is fine for narrow task-specific problems, but is not really the end-all of AI, and does not even seem to work well on all well-defined tasks. As an example of another kind of inference, how about mathematical reasoning? I purposely pick one here that is seemingly very formal; should be possible for a computer to do it. Mathematicians are somehow able to invent conjectures and prove theorems without first being exposed to terabytes of labeled mathematical facts. The scientific method is kind of an even messier version of this. Or to take something laypeople do, people can usually learn games to at least a passable level from just a handful of playthroughs, not AlphaGo-style millions of plays (imagine if you had to play even 1,000 games of MT:G before you got the basic hang of it...).

All this kind of stuff is quite well-represented in the literature though, if you mean the scientific literature. Pop-press AI writing tends to cover a pretty specific subset of what's going on at AI conferences.

Reply on news.ycombinator.com