Let's consider only "safe for work" Internet content:
Google/Bing have done well with keyword/phrase searching with results sorted by popularity and date.
Ng seems to be thinking that the big change will be having speech and images coming from the users as their input to the search process. My guess is that this will not be very important. I also guess that search for Internet content based on speech and images will become more important.
Yes, likely Ng can get a lot of pictures of what he knows to be, say, Ferraris and use some of them as the training set with a neural network to
identify Ferraris and test the training with the rest of the pictures. Okay. Maybe his neural network will be able to identify Ferraris. So, he could repeat this training for, say, 100,000 objects -- Fords, bread, airplanes, jewelery, Victorian houses, .... Maybe there will be some value there.
My view is that the future of Internet search is quite different.
words are an incredibly efficient mechanism for communicating with google. Google image search is great for some things, like you find a picture of a sculpture on tumblr and it's not credited, use the photo as the input to find the source. But Taking photos of a ferrari seems like a dreadfully inefficient way to make a search. Maybe identifying flowers would be a better example, but most cars have in very clear letters the make and model written on the back.
As I understand it, you have described what Google does with words and pictures well.
Ng wants neural networks to identify things, maybe Ferraris.
Then he wants a search user to send a picture as their input for the search they want to do. So, then the user might be able to find more pictures of Ferraris. Maybe.
For what Ng is doing, I doubt that flowers would work because there are far too many too different cases of flowers.
Search by keywords/phrases by Google/Bing has worked very well for a huge collection of Internet content.
But I am guessing that in a sense Ng is correct about images and sounds -- there stands to be a lot more such content on the Internet in the future.
Generally I'm guessing that there is also a huge collection of Internet content, searches people want to do, and results they want to find where search via keywords/phrases such as via Google/Bing is from poor down to useless. Thus, my guess is that a new means of search is needed.
For what Ng is doing, I doubt that flowers would work because there are far too many too different cases of flowers.
Only if you see it as a supervised learning problem. An alternative is to find the nearest matches, after which you can let human intelligence make the final visual match. Often, the webpage containing the matched image will have enough context to identify the object being searched for.
Are you suggesting that there is a distinction between "searches people want to do" and "results they want to find" (besides the purely functional one of search -> result). e.g. there are search results people want, even though they don't know they want them? Unknown unknowns that people will retrospectively be grateful for?
I'm guessing the work flow would be something like "I want intermittent rotation" except probably not expressed as literately. Maybe a sketch or a lot of babble along those terms gets searched on, whatever input format. If it works, maybe you find a Geneva mechanism.
If you have enough domain experience to search for "continuous rotation rotary intermittent rotary" then you'll find it, but if you are mechanically illiterate you may not know what intermittent means or rotary... maybe.
It would be a truly amazing display of AI to be given a really poor sketch of a Geneva mechanism, it'll find a really nice blueprint. I'd be impressed if this found the hypoid gear in a differential. I'd be more impressed if someone who doesn't understand the concept or reason for a torsen differential was able to none the less search it.
As a concrete example theres a pretty impressive lego torsen(-ish) diff out there. Its easy to find if you google for the terms. I'd be impressed if you could give a sketch to a search engine and find this lego diff.
I'm not trying to compete with Google/Bing where their work
with keywords/phrases, page rank popularity, and, say, date, work well. And for a huge pile of Internet content, search, and results, they do work well.
For the searches you mention, I believe that maybe for one of them it could be possible to improve on Google/Bing, but I don't believe that real AI would be needed. For describing how to build such a search engine, that might take more than the 10,000 character limit on HN posts!
Yes, there can be some differences. E.g., a user might search for something that does not exist. So, they can do the search, that is, attempt, but not find the results.
Or, there's a lot of content on the Internet; a lot of it, a lot of people want; to get it they want to do searches that promise to be able to find that content. Nothing more obscure than that.
Comments
Let's consider only "safe for work" Internet content:
Google/Bing have done well with keyword/phrase searching with results sorted by popularity and date.
Ng seems to be thinking that the big change will be having speech and images coming from the users as their input to the search process. My guess is that this will not be very important. I also guess that search for Internet content based on speech and images will become more important.
Yes, likely Ng can get a lot of pictures of what he knows to be, say, Ferraris and use some of them as the training set with a neural network to identify Ferraris and test the training with the rest of the pictures. Okay. Maybe his neural network will be able to identify Ferraris. So, he could repeat this training for, say, 100,000 objects -- Fords, bread, airplanes, jewelery, Victorian houses, .... Maybe there will be some value there.
My view is that the future of Internet search is quite different.
words are an incredibly efficient mechanism for communicating with google. Google image search is great for some things, like you find a picture of a sculpture on tumblr and it's not credited, use the photo as the input to find the source. But Taking photos of a ferrari seems like a dreadfully inefficient way to make a search. Maybe identifying flowers would be a better example, but most cars have in very clear letters the make and model written on the back.
As I understand it, you have described what Google does with words and pictures well.
Ng wants neural networks to identify things, maybe Ferraris. Then he wants a search user to send a picture as their input for the search they want to do. So, then the user might be able to find more pictures of Ferraris. Maybe.
For what Ng is doing, I doubt that flowers would work because there are far too many too different cases of flowers.
Search by keywords/phrases by Google/Bing has worked very well for a huge collection of Internet content.
But I am guessing that in a sense Ng is correct about images and sounds -- there stands to be a lot more such content on the Internet in the future.
Generally I'm guessing that there is also a huge collection of Internet content, searches people want to do, and results they want to find where search via keywords/phrases such as via Google/Bing is from poor down to useless. Thus, my guess is that a new means of search is needed.
What do you think?
Only if you see it as a supervised learning problem. An alternative is to find the nearest matches, after which you can let human intelligence make the final visual match. Often, the webpage containing the matched image will have enough context to identify the object being searched for.
What is a near match is an issue.
Are you suggesting that there is a distinction between "searches people want to do" and "results they want to find" (besides the purely functional one of search -> result). e.g. there are search results people want, even though they don't know they want them? Unknown unknowns that people will retrospectively be grateful for?
I'm guessing the work flow would be something like "I want intermittent rotation" except probably not expressed as literately. Maybe a sketch or a lot of babble along those terms gets searched on, whatever input format. If it works, maybe you find a Geneva mechanism.
If you have enough domain experience to search for "continuous rotation rotary intermittent rotary" then you'll find it, but if you are mechanically illiterate you may not know what intermittent means or rotary... maybe.
It would be a truly amazing display of AI to be given a really poor sketch of a Geneva mechanism, it'll find a really nice blueprint. I'd be impressed if this found the hypoid gear in a differential. I'd be more impressed if someone who doesn't understand the concept or reason for a torsen differential was able to none the less search it.
As a concrete example theres a pretty impressive lego torsen(-ish) diff out there. Its easy to find if you google for the terms. I'd be impressed if you could give a sketch to a search engine and find this lego diff.
I'm not trying to compete with Google/Bing where their work with keywords/phrases, page rank popularity, and, say, date, work well. And for a huge pile of Internet content, search, and results, they do work well.
For the searches you mention, I believe that maybe for one of them it could be possible to improve on Google/Bing, but I don't believe that real AI would be needed. For describing how to build such a search engine, that might take more than the 10,000 character limit on HN posts!
Yes, there can be some differences. E.g., a user might search for something that does not exist. So, they can do the search, that is, attempt, but not find the results.
Or, there's a lot of content on the Internet; a lot of it, a lot of people want; to get it they want to do searches that promise to be able to find that content. Nothing more obscure than that.