Neural networks

By joining our community you will have the ability to post topics and access other forums reserved for members. Registration is quick, simple and absolutely free. Join our community today by clicking here.

Posted: Tue May 09, 2017 10:34 am

So, you all have probably read sometime, somewhere (at the very least) of the hot new thing of nerual networks and self-learning algorithms.

Such networks have been shown to produce

• Poetry http://neuralnetpoetry.blogspot.com/
• Music https://www.youtube.com/watch?v=SacogDL_4JU
• Art http://www.smithsonianmag.com/smart-news/new-rembrandt-created-347-years-after-the-dutch-masters-death-180958664/

These are, but examples of how neural networks can learn from samples. The greater the sample, the better the ability to learn. Rather than providing an 'average' from the samples, these create new and unique content based on patterns they've learned by looking at the samples.

It is truly fascinating technology and I recommend checking out those links.

Today, however, I found another interesting use of this technology. Especially in regards to our baiting world, where both sides, that is lads and baiters, take different personalities and use them to their favour.

https://lyrebird.ai/demo - listen to the first demo. Had I not known, I would likely have been convinced! Great stuff.

It also shows that soon, as far as anything digital goes, you will not be able to trust anything. With this, and all thinks CGI, photoshop, post-processing and what-not, the line between reality and a good fake will grow thinner and thinner. I'm not saying lads will have this sort of technology available anytime soon, and considering their ever so terrible photoshops, there's no need to be afraid they'd be able to fully impersonate Donald Trump for that trunkbox messages next week. (altough Lyrebird wants to make sure anyone has free and open access to it) - but it's an interesting topic to think about and discuss nonetheless.

Where do you see this technology in the next 5, 10, 20 years? And will there, for example, be such a time when you will be able to go see Casablanca 4K remake done entirely with the original cast and you wouldn't even notice it has been generated by computer? And the possibilities really go on, of course.

Posted: Tue May 09, 2017 11:18 am

Yes for sure. We have already seen advances based on various flavours of deep learning and variations thereof, maybe most relevant to baiting is the very rapid improvements in machine translation that has already allowed them to target a much wider audience and will continue to improve further.

This has been most evident in text to text translation, but speech to speech is very rapidly advancing. So far there are not very good training samples that would allow this to work well for typical lad accents but the time is coming. For me projects like lyrebird are not such a worry in that famous people's speech profile can be imitated. But rather, with a very small sample an ordinary person's speech might be sufficiently imitated, and used in scams such as the "mugged on holiday" or phishing upon people who know them.

I find it especially interesting that these models develop representations and processes that do not appear to be "human like", yet are still extremely effective at solving problems they are applied to ( thinking of the recent Go developments here). The "unknowns" are creating some concerns about just what may be going on -- very weird patterns of failure such as when you enter in a long series of the same word to Google translate.

Posted: Tue May 09, 2017 2:43 pm

On a similar vain of technology, its quite amazing to see how far AI visual recognition has become accessible to the regular end user. If you have android, try Allo take a 3/4 view picture of your car and send it to friend on Allo. Allo will not only recognize the make of the car, sometimes even the model. in its reply suggestions.

What this has to do with baiting? Imagine this applied to your inbox, you get an attachment from a lad. Gmail warns the picture has been manipulated. You could argue that it would teach the lads to 'shop better, but imagine the level of skill you would have to be to fool a computer that anylaizes the picture per pixel and does this millions of times a day.

Posted: Tue May 09, 2017 5:00 pm

Interesting stuff for sure. I read a research paper many moons ago where the goal was to get the AI to create harmony on the fly when given a melody. The sample set was all Bach and then an algorithm was built to help the thing learn.

If I remember correctly it was a total failure. There was too much variation and "feel" to the music and the AI could never really make it work.

Which I found made me kinda glad...

Hello I'm New here! Joined: 31 Aug 2016 Posts: 17

Pardon me to interrupt. Back in the day it was all about gate doping (micro tech), and now it's neural bio stuff. Neural networks, and more carbon tubes, etc. Buckyballs?

Where is my nematode...