Crosswords1 min ago
Oh Dear Another Dead Road User Killed By Computer
129 Answers
http://
When are we going to accept that our current software ability is insufficient for this application? at least Uber has the sense to halt their tests.
When are we going to accept that our current software ability is insufficient for this application? at least Uber has the sense to halt their tests.
Answers
If this stuff Is that good, why are we about to spend billions on infrastructu
22:28 Tue 20th Mar 2018
From one of my links and written by a coder:
‘Some of the most significant advances in AI over the past decades involve “neural networks” and variations of this idea, including those recently branded “deep learning”. Those are computational paradigms inspired by the structure of the brain. In this sense, yes, our understanding of basic functions of the building blocks of the brain have helped us make progress with machine learning.’
Where do computational paradigms fit in with your ‘it’s just more power’ claim?
‘Some of the most significant advances in AI over the past decades involve “neural networks” and variations of this idea, including those recently branded “deep learning”. Those are computational paradigms inspired by the structure of the brain. In this sense, yes, our understanding of basic functions of the building blocks of the brain have helped us make progress with machine learning.’
Where do computational paradigms fit in with your ‘it’s just more power’ claim?
There's nothing remotely futuristic about driverlesss trains, they've been around forever with the Queen famously riding the Docklands Light Railway in 1987. The Post Office had driverless underground railways in London from I think the 1920's....most people that have been to a fairground have been on a driverless train.
But they run at relatively low speed, directional control isn't a problem and they tend not to meet other traffic - a far cry from motor vehicles which must be able to cope with the M25, cart tracks and everything in between, in all weathers day and night.
But they run at relatively low speed, directional control isn't a problem and they tend not to meet other traffic - a far cry from motor vehicles which must be able to cope with the M25, cart tracks and everything in between, in all weathers day and night.
I think it's safe to say that the AlphaZero software isn't just about "more power to process more" -- end of, or otherwise. Indeed, the point (made in the recent paper) was that, if anything, there was less computing power thrown at it ultimately (once the neural network had been trained for long enough, at least), as compared to equivalent programs in chess such as Stockfish that do rely in large part on computing power.
I don't think AlphaZero or equivalent is yet at the stage where it's capable of being used in any meaningful time on a single machine, but then it's still an active research project and in the long run it's possible, or even probable, that there will come a point where similar, flexible, software is ready to solve many such problems without just having stupidly fast processors.
Finally, on the subject of Go, the lesser-known program "LeelaZero" was written towards the end of last year, and doesn't rely on TPUs to train -- instead making use of multiple home computers to achieve the same effect. It took a little longer (a few bugs) but once again the effect is remarkable and after about a month the program was well beyond most strong amateurs even with only a couple of thousand simulations per move.
All of this is by the by, but the point is that AI and machine learning are indeed progressing beyond the "old days" -- and I am surprised, to say the least, that you are either unaware of this or rather arrogantly dismissive of said progress. One might almost wonder if you think that the only difference between computing now and computing when you started was that we've stopped using punch cards.
I don't think AlphaZero or equivalent is yet at the stage where it's capable of being used in any meaningful time on a single machine, but then it's still an active research project and in the long run it's possible, or even probable, that there will come a point where similar, flexible, software is ready to solve many such problems without just having stupidly fast processors.
Finally, on the subject of Go, the lesser-known program "LeelaZero" was written towards the end of last year, and doesn't rely on TPUs to train -- instead making use of multiple home computers to achieve the same effect. It took a little longer (a few bugs) but once again the effect is remarkable and after about a month the program was well beyond most strong amateurs even with only a couple of thousand simulations per move.
All of this is by the by, but the point is that AI and machine learning are indeed progressing beyond the "old days" -- and I am surprised, to say the least, that you are either unaware of this or rather arrogantly dismissive of said progress. One might almost wonder if you think that the only difference between computing now and computing when you started was that we've stopped using punch cards.
‘I can answer but I don't want to derail my thread’
As I’ve pointed out previously, your assertions that driverless cars are a long way from reality is based on your spurious assertions that AI doesn’t exist or at least cannot be applied to this technology, therefore discussing what ‘thinking’ is, is pretty crucial to moving forward on this debate.
As I’ve pointed out previously, your assertions that driverless cars are a long way from reality is based on your spurious assertions that AI doesn’t exist or at least cannot be applied to this technology, therefore discussing what ‘thinking’ is, is pretty crucial to moving forward on this debate.
I think what ZM is -- and, for that matter, I am -- asking is quite pertinent to this thread. Has AI research advanced in the last 50 years, and if so how do you see the recent research by DeepMind fitting in to that picture? Or, if not, why not?
All of this clearly *does* impact on the potential for driverless cars, since, inevitably, they will have to rely on neural networks, deep learning and related architecture as part of their software.
All of this clearly *does* impact on the potential for driverless cars, since, inevitably, they will have to rely on neural networks, deep learning and related architecture as part of their software.
As I said above, playing GO or chess, ie fixed rules closed system are not comparable to the infinite conditions, rules, real and derived, thinking, instant dismissal of options that thought has. I see those developments as useful but never adequate enough to cope with driving, possibly railways in the medium term. I hope for some software breakthrough that takes us to the next level. Only recently did we acquire enough processing power to defeat humans at Chess and GO, the reason for that is that, flawed humans may be, but they can cut out huge swathes of nonsense instantly where as a computer has to work through it all hence processing power is needed. Yes the algorithm will "learn" by storing and using that later but basically it's using data and processing power, not thought. Perhaps if you can try and read what I have said on this thread rather than just trying to insult me because a dare to go against the "fashion" you may see something you have previously missed.
I hope you won't take this the wrong way but... I don't come to Answerbank to learn about the latest trends in AI. Even if it's you talking about them.
Still, the comment about closed v. open systems is pertinent. On the other hand, that's still a massive shift from just the start of this page, when you proudly and utterly incorrectly announced that "software hasn't really changed since the early days".
I suspect that there's still a lot people can learn from what you've posted. Just not much about AI.
Still, the comment about closed v. open systems is pertinent. On the other hand, that's still a massive shift from just the start of this page, when you proudly and utterly incorrectly announced that "software hasn't really changed since the early days".
I suspect that there's still a lot people can learn from what you've posted. Just not much about AI.
Latest Posts
ChatterBank1 min ago
This 'Bad Actor ' Trying To Harm Reform.
News2 mins ago
Crosswords3 mins ago
Who's Got The Guts To Take This...
News4 mins ago
ChatterBank9 mins ago
Related Questions
Sorry, we can't find any related questions. Try using the search bar at the top of the page to search for some keywords, or choose a topic and submit your own question.