Do you think artificial intelligence and DFT codes can be merged?

[proteo]

Hello! Do you think artificial intelligence and DFT codes can be merged? Is it time for the intelligent DFT?

 

[Grinkle]

Do you mean Design For Test?

 

[proteo]

I mean a scientist (assistant) AI. If there are gamer AI's (chess, go vs.), there should be scientist AI. Google allows people to use its codes. A good programmer (maybe I) can add siesta packages in AI code. Maybe, we say to google assistant "please, calculate these materials' electrical band structure". Maybe this conservation will occur next year.

 

[jim mcnamara]

We do not want speculation in the science forums.
Moving to General Discussion.

 

[jedishrfu]

We are a long way from having a scientists assistant as you’ve described. The state of AI today are basically some clever search space algorithms or neural nets working with big data to create the illusion of intelligence. What you describing is the Star Trek Computer that Spock would converse with.

When I watched the old shows recently I realized that Spock’s dialog was a lot like a conversational SQL statement. He would say something like

Computer, take this sensor data and correlate with data collected from the previous mission and order it by planetary alignment. It had to sound scientific while at the same time something a computer might do. The computers female voice would respond Working and then provide the answer.​

In any event, this human computer dialog is closer to what we still have today



You could research efforts on AGI or ASI which is where AI research is heading. Perhaps the confluence of the Quantum Computer, Huge Data and a better understanding of how intelligence emerges from our brains will happen and we will reap the benefits of and unanticipated side effects of an artificially intelligent mind.

 

[Klystron]

I mean a scientist (assistant) AI. If there are gamer AI's (chess, go vs.), there should be scientist AI. Google allows people to use its codes. A good programmer (maybe I) can add siesta packages in AI code. Maybe, we say to google assistant "please, calculate these materials' electrical band structure". Maybe this conservation will occur next year.

Much assistance has already been automated and is in regular use. For example Archimedes or John von Neumann might go to lunch while asking an assistant to bring a selection of scrolls/books from the library required for his work. Now this information can be retrieved in milliseconds with roughly the same commands to a computer. Many other tasks are regularly performed by our inventions but we may not think of it as 'artificial intelligence' but simply commonplace technology.

With enough funding much is possible. For example a laboratory might receive packaged sample containers, open packages, open containers, place samples in appropriate device, examine samples, run different experiments and report results with nary a human in sight. For various reasons one company or lab probably analyses DNA, another reactor fuel pellets, and another examines fecal swabs for cancer cells but the machines do not care. Also, no humans in the loop means no voice commands are required unless the scientist receiving the final reports prefers spoken language.

 

[proteo]

But none of them DFT. Voice command may not be required for DFT but it is first step for AI understanding science. Genetic programming was able to find Kepler's Rules on its own. As I said before my aim is INTELLIGENT DFT.

 

[proteo]

You can discuss your results with AI. It is pure science in both way.

 

[DennisN]

A good programmer (maybe I) can add siesta packages in AI code. Maybe, we say to google assistant "please, calculate these materials' electrical band structure". Maybe this conservation will occur next year.

Here is a pretty good short video I posted from a couple of years ago which describes the remarkable difficulties involved (and Natural Langauge Understanding in general is notoriously difficult):



See also these previous threads on PF:

• Asking bots for help (see e.g. my post #6 and others)
• What makes something artificially intelligent? (see e.g. my post #4)

Nevertheless, it is a very interesting branch of computer science, and probably my favorite one. But it is also very difficult, which I know from experience .

Edit:
I'd like to add that I believe you also mentioned speech recognition, which in itself is different from natural language understanding. Speech recognition has "simply" the objective of taking audio and trying to convert it into text. But after that, that's when the really, really difficult things start , which is trying to get a machine to successfully parse and artificially cognitively understand what is actually being meant in the text.

 

[proteo]

Of course there are difficultes, but AI could interprent emprical results and with support of the quantum computing, comments of AI might be interesting.
https://www.machinegurning.com/rstats/kepler_gp/
https://jontse.com/courses/files/cornell/cs5722/11-11-18_GeneticProgramming.pdf
https://www.kaistaats.com/blog/2015/09/genetic-programming-keplers-law/
https://books.google.com.tr/books?id=Bhtxo60BV0EC&pg=PA258&lpg=PA258&dq=genetic+programming+kepler's+law&source=bl&ots=9qfPkqb0LM&sig=ACfU3U0v2UqTa96gw94NvXzhZbYD241-vA&hl=tr&sa=X&ved=2ahUKEwjai8KwhYnhAhU2wsQBHT0VC4gQ6AEwBXoECDMQAQ#v=onepage&q=genetic programming kepler's law&f=false
There is huge but only one step to discuss scientific problems with AI simultaneously.

 

[proteo]

Dear editors, my thread is not different from this: https://www.physicsforums.com/threads/limits-of-machine-learning.921251/ Why did you move my thread under general discussion?

 

[proteo]

Here is a pretty good short video I posted from a couple of years ago which describes the remarkable difficulties involved (and Natural Langauge Understanding in general is notoriously difficult):



See also these previous threads on PF:

• Asking bots for help (see e.g. my post #6 and others)
• What makes something artificially intelligent? (see e.g. my post #4)

Nevertheless, it is a very interesting branch of computer science, and probably my favorite one. But it is also very difficult, which I know from experience .

Edit:
I'd like to add that I believe you also mentioned speech recognition, which in itself is different from natural language understanding. Speech recognition has "simply" the objective of taking audio and trying to convert it into text. But after that, that's when the really, really difficult things start , which is trying to get a machine to successfully parse and artificially cognitively understand what is actually being meant in the text.

Thank you for video. I worked on image processing algorithms and genetic algorithm.

 

[DennisN]

Thank you for video. I worked on image processing algorithms and genetic algorithm.

Nice! Those are fun AI technologies.

I specialized in AI at university, particularly neural networks, but I studied other AI things too, and did some of my own things both in my free time and a bit at work. One thing I remember from my time at university was how unaware I was of how difficult it is for machines to interpret human language; I'd say it is very hard to imagine just how difficult it is until you have tried implementing it;

One of our projects at university was to implement a QAS (question answering system) for airports, where the user, typically a passenger, would manually enter a question in natural language regarding flights, arrivals, departures, times, dates etc. Now, since this was only in the context of airports this problem was what is called a closed domain problem, which makes it much easier (but still difficult!). We managed to implement it with pretty good results. I don't remember which programming language we used, and I got curious, so I will dig up our old report and have a look at it, so I may post about it again in this thread later.

This can be compared to problems with larger domains; let's say we add the domains of bus terminals, ship terminals, train stations, and you may understand that the complexity goes up quickly, since the system in this case would have to be able to handle different physical contexts.

And then we have the ultimate domain: the open domain, where the context can be anything. This is the ultimate goal, the holy grail of AI natural language understanding, and the absolutely most difficult task. Computer science is not even close to anything like it at all at the moment.

Edit:

IBM's Watson computer is as far as I know one of the most advanced language systems, but even though it is sort of, or rather a tiny bit open domain, it is still not the real deal, since it is "only" capable of answering (in this case, Jeopardy) questions one and one. But it is a remarkable achievement!

The real deal, the holy grail, is to be able to successfully have two-way conversations with a machine, regardless of context, and where the machine is able to follow a two-way conversation from beginning to end, which means it must be able to handle time contexts too (along with many other different contexts ), which means it also must have some sort of advanced adaptive memory.

 

[DennisN]

Hi again, @arcedia,

I don't remember which programming language we used, and I got curious, so I will dig up our old report and have a look at it, so I may post about it again in this thread later.

I found our old report and we used Scheme as a programming language (which is a dialect of Lisp). It was a very long time since I programmed Scheme, but I remember I really liked it for particular AI purposes.

I got a bit nostalgic so I scanned our old report, and I share here the first page (including the abstract) and also an example of actual results of a Natural Langauge Understanding (NLU) session with the question answering system (QAS);

Page 1 (with abstract), which basically describes the general problem regarding NLU and QAS:

The objective in our case was that the system was supposed to be able to answer questions stated in arbitrary natural language from a flight passenger who in some way stated where he/she is and where he/she wants to go, and optionally also stated information about when he/she wants to depart or arrive. Then the system should reply with different available flights and departure and destination times to different airports. If I remember correctly, the system was dynamic and could in principle handle any number of flights and cities.)

And here is an example of entered questions by the user (in italics) and followed by responses by the system:

 

[gleem]

This is one of those threads that piques my curiosity and sends me out looking for answers. In particular where are we in natural language processing. Well I haven't found the right article yet but it did find and article on a website of particular interest.titled "WE SUMMARIZED 14 NLP RESEARCH BREAKTHROUGHS YOU CAN APPLY TO YOUR BUSINESS". The interesting point is that for each of the 14 articles it gives

The original abstract
The sites summary of the article
The core idea of the paper
The key achievement of the research
The opinion of the community
Possible future research
Possible business applications
Availability of the code.

This should be of interest to those who are watching the development of NLP and AI in general.

 

[DennisN]

This should be of interest to those who are watching the development of NLP and AI in general.

Thanks, that looks interesting, I'm bookmarking it, and I'm going to check it out!