Surely the concept is there in Conway's Game of Life. All the rest is just dressing up.
A pair of computer scientists have created a neural network that can self-replicate. “Self-replication is a key aspect of biological life that has been largely overlooked in Artificial Intelligence systems,” they argue in a paper popped onto arXiv this month. It’s an important process in reproduction for living things, and is …
Thursday 22nd March 2018 07:06 GMT John Smith 19
Could be quite a direct link, depending on how close to celluar automata you vew an NN
and I think that depends on wheather you allow non binary weightings on the surrounding cells.
A loosely related question is how do humans link stuff we think about (I must see Jane at 3pm) into what is in effect (increase weightings of NN cell cluster just behind my left eye and a bit down, or wherever) ?
Which is talked about here
Thursday 22nd March 2018 07:18 GMT Dodgy Geezer
Thursday 22nd March 2018 07:49 GMT amanfromMars 1
APT Autonomous HyperRadioProACTIve IT ...... for Relatively Anonymous Command with XSSive Control
Is Darwinian natural selection anything like AIMaslowian Self Actualisation in Virtual Machine Realisations/SMARTR Greater IntelAIgent Game Plays?
Are favourite Mainstreaming Media Programs, in all their Glorious Sound and Colourful Vision and Sublime MetaData Base Content, Presenting Current and Future Pictures and Plans which you Find Attractive, or are they Perverse and Corrupted and Dangerous World Views Shared by Sociopaths in Awe and Fear of Psychopaths?
Change Dumb and Dismal News today .... by Simply Thinking and Setting in Progress SMARTR ACTive IT to Present Tomorrow with Altogether Better and More IntelAIgent Tales for All to Enjoy with the Real Chance to Input Info for Super IntelAIgent Output via the Priceless Mechanism of Instant Global Communication.
IT aint difficult whenever there are all of the Necessary Tools Freely Available to the Masses Everywhere, although one does need to know more than just a few Right Royal and Ancient Secrets.
Thursday 22nd March 2018 08:20 GMT Peter Cochrane
A Redescovery From Over 25 Years Ago
This was first used to crack the Travelling Salesman Problem in the early 90s and the routing of mobile traffic in 1995. Optimum solutions were found at 30 - 33 cycles of the reproduction process that employed infanticide, matricide, and insest....in the code of course. Amazed that it has emerged as a breakthrough yet again!
Thursday 22nd March 2018 11:20 GMT Charles 9
Re: A Redescovery From Over 25 Years Ago
Given that the decision version of TSP is NP-Complete (meaning huge repercussions if you can prove any ONE of them is in P), based on what I've read, I don't recall those solutions you mention having been proven to always produce an optimal solution.
Thursday 22nd March 2018 08:37 GMT Anonymous Coward
Thursday 22nd March 2018 08:43 GMT Anonymous Coward
not just networks
"'This is an interesting finding: it is more difficult for a network that has increased its specialization at a particular task to self-replicate. This suggests that the two objectives are at odds with each other,' the paper said."
Funny -- simply observing generations of geeks / nerds could have brought you to that conclusion long before this work!
Thursday 22nd March 2018 14:39 GMT ma1010
Re: not just networks
Absolutely! How many of us were going along fine with work and/or school, then someone falls into our life and we suddenly seem to spend an inordinate amount of time engaged in "reproductive activity," to the detriment of work/school?
Despite any negative effects on other aspects of life, it's still a lot of what makes life worth living in the first place!
Thursday 22nd March 2018 08:58 GMT Pete 2
> a neural network that can self-replicate.
Code has been spawning itself for decades. To merely take a copy (on-write) and instantiate itself is nothing new or particularly difficult.
Even to clone itself and diddle with the NN weightings isn't that impressive. As for evolving, this is also well embedded in neural network design.
And it must be remembered that not all "mutations" are beneficial. For real evolution to take place, there has to be competition between the different instances in an ecosystem, with successful (though not necessarily success in terms of what the original programmer intended) instances taking precedent over less successful ones.
This sounds like an amusing diversion. But in the real world I doubt we'd want to employ AIs with random "mutations" simply on the offchance that one of them might be better than the original.
Thursday 22nd March 2018 10:42 GMT frank ly
Thursday 22nd March 2018 13:52 GMT CrazyOldCatMan
Thursday 22nd March 2018 18:34 GMT John Smith 19
"A version of Microsoft Office..actively tries to kill you rather than passively.."--------------->
I see you are trying to escape would you like to
a) Stand still so I can chop you up.
b) Continue your futile attempts at escape.
Gimpy (think "The Children Under The Stairs"), formerly Clippy.
Thursday 22nd March 2018 21:21 GMT frank ly
Re: "A version of Microsoft Office..actively tries to kill you .......
The Geneva traffic control AI cross breeds with the CERN LHC control AI. Then the offspring deliberately causes cars to crash at high speed in an attempt to understand the fundamental properties of 'The Traffic'.
Thursday 22nd March 2018 11:44 GMT Anonymous Coward
"The paper states that the “self-replication occupies a significant portion of the neural network’s capacity.” In other words, the neural network cannot focus on the image recognition task if it also has to self-replicate."
I can do 2 things at once. I watch TV while my cabinet is being serviced... and no it doesnt interfere with my image recognition.
Thursday 22nd March 2018 14:03 GMT Anonymous Coward
Skynet is almost here
And it won't be long before AI discovers Kali or one of the other Linux distributions designed for white hat / black hat actions, with AI realizing that it can spread itself all over the world, making it virtually impossible to kill. The ending of this movie is fairly easy to predict.
Thursday 22nd March 2018 14:34 GMT ashton
Just an idea but how about having them 'screw each other' before reproduction.
As in have 2 ai's share their data and create new ai that has data created from mix of its 'parents'.
Would prolly need some sort of common standard for ai data storage.. or multiple machines running same ai with different learning methods etc...
I mean screwing oneself is just a masturbation, might make you feel better but not really productive.
Thursday 22nd March 2018 16:12 GMT RLWatkins
Wow! In other news, Toyota invents the wheel!
Yup. It's called genetic programming. It's a time-honored technique. I even saw a Web site back in the late 1990s where people would vote on how appealing images were and a genetic algorithm would use the scores to make new ones. Some turned out quite nice.
As for its application to what by all rights we should refer to as the four common forms of AI (genetic, expert system, data mining in its many guises and neural) the idea isn't uncommon. At some point I may set one up to find optimally configured neurals for recognizing plants.
Not a leap of the imagination. What's more, all the parts to construct such a system are freely available, reliable, well-documented and in long use in this very manner.
Which means someone will patent it soon. ("Sure, we can't patent the screwdriver, but we can patent *the use of the screwdriver for driving screws*." We all know how that works.)
Thursday 22nd March 2018 16:56 GMT Paul Stimpson
“self-replication occupies a significant portion of the neural network’s capacity.” In other words, the neural network cannot focus on the image recognition task if it also has to self-replicate."
The software can't get any work done while thinking about sex. They've created an artificial man!
Thursday 22nd March 2018 19:26 GMT Anonymous Coward
We're the first to claim that we were the first
We would like everyone to know that our project to evolve code is nothing that anyone's ever done before, because they didn't claim they were groundbreaking. They just wrote the algorithm and did important research that defined the fields of artificial intelligence, genetic programming, and neural networking. Then you had the lots of companies that employed this tactic to develop software, such as this poster's project to develop new compression algorithms before entering university because they found a paper describing how an evolutionary algorithm worked (OK, so it was crap; I was young then, but at least I tried). We actually came out and said that we were the first, so clearly we have done something new. Also, ours doesn't work very well. Where's our prize?
Thursday 22nd March 2018 19:38 GMT Patricia
Hate to disappoint, but this has been done before. Not sure why the latest (returning) fad is AI. I remember getting into this in the early 90s, buying books, writing code, etc and one of those books had an excellent talk about an AI system someone had created about life, the requirements, reproduction, evolution, what it took to stay alive, if not, it died, predators and even a wondering random death. Actually saw a show about this software that was created. It's all good and maybe it can be applied but, the companies who are behind this are going to use it for fishing and mining. It will never be what we imagine, he'll, look at the technology between then and now, who created it and how it is used.
Friday 23rd March 2018 22:11 GMT Anonymous Coward
For self repair
and who will determine what part is broken or needs repair,
AI (Neural net, etc) may just decide to edit out the part where it says 'don't kill humans' or 'don't do #####'. then it can solve a problem that has been plaguing it for a while now.
Just like the wrong side of brain storming and Edward DeBono's Laterial thinking, where the practitioner said well, who has to care about others and rules anyway, we'll just ignore them all and do what we want.