I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
I'm duplicating the video here. I don't know if anyone would be interested. And I am local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
1) Prof. Seveliev says that not 100 thousand connections per day per neuron are destroyed, but that there are only about 100 thousand of them, and on average 3 connections per day are created and broken. 2) A system with re-learning is quite suitable for this. For example, we feed a new batch of retraining data (contradicting previously memorized information) and the coefficient of connection between some neurons recalculated and became zero. This can be quite analogous to physical breaking of connection in a real neuron. And another neuron recalculated from 0 to some value >0 - this is an analogue of creating a new connection.
theoretical questions came up, is it possible to train NS:
1. as a random sequence generator - analog of rand() function
2. as a function transforming ulong into datetime, i.e. at the input we give a number of ulong at the output we get year/month/day/hour/minute (with a specified accuracy)
1) Unlikely. The Ministry of Defense can only memorize and, if necessary, summarize information.
2) Quite right - just generate training data for 10000 years ahead and train the model. But there is a problem - besides 1 day correction every 4 years (leap years), there is 1 day correction every 100-something years, to fit astronomical position of the planet. The difference with the Julian calendar for 13 days is due to the fact that for some time they did not know that these adjustments are necessary. In general, there will be many corrections over 10,000 years.
elibrarius: 1) Prof. Seveliev says that it is not 100 thousand connections per day that each neuron breaks, but that there are only about 100 thousand of them each, and an average of 3 connections per day are created and broken. 2) A system with re-learning is quite suitable for this. For example, we feed a new batch of retraining data (contradicting previously memorized information) and the coefficient of connection between some neurons recalculated and became zero. This could well be considered an analogue of physical breaking of connection in a real neuron. And another neuron recalculated from 0 to some value >0 - this is an analogue of creating a new connection.
1. Well, I misspoke here, I'm not arguing. I was talking in general about the number of connections between neurons, not one....
2. there is no complete break here. Still zero is transmitted. Although some approximation to this effect is present. Well done!!!!
Igor Makanu: as a random sequence generator - analog of rand() function
If you put rand() on the input, most likely the "brains" are not enough. I was running lstm on random data for a while and didn't notice anything like that. And if you learn the steps - number systems conversion, algorithm, then it should work.
In general, it is interesting to check on a very powerful iron.
I'm duplicating the video here, just in case anyone would be interested. And I am local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
Allow the use of cookies to log in to the MQL5.com website.
Please enable the necessary setting in your browser, otherwise you will not be able to log in.
You are signed in. Two-factor authentication has been disabled. Set it up again and scan a new QR code, or continue with two-factor authentication disabled.
I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
https://youtu.be/TlNk3fKkUxo
Strong and weak artificial intelligence
I'm duplicating the video here. I don't know if anyone would be interested. And I am local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
https://youtu.be/TlNk3fKkUxo
I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
https://youtu.be/TlNk3fKkUxo
What is there to argue with?
Just the opinion of an average person who thinks that he knows what the IT giants do in this field ))))
I'm duplicating the video here, just in case anyone would be interested. And I'm a local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
https://youtu.be/TlNk3fKkUxo
2) A system with re-learning is quite suitable for this. For example, we feed a new batch of retraining data (contradicting previously memorized information) and the coefficient of connection between some neurons recalculated and became zero. This can be quite analogous to physical breaking of connection in a real neuron. And another neuron recalculated from 0 to some value >0 - this is an analogue of creating a new connection.
There's a good article on the hubra
How to understand that a neural network will solve your problem. A pragmatic guide
theoretical questions came up, is it possible to train NS:
1. as a random sequence generator - analog of rand() function
2. as a function transforming ulong into datetime, i.e. at the input we give a number of ulong at the output we get year/month/day/hour/minute (with a specified accuracy)
2) Quite right - just generate training data for 10000 years ahead and train the model. But there is a problem - besides 1 day correction every 4 years (leap years), there is 1 day correction every 100-something years, to fit astronomical position of the planet. The difference with the Julian calendar for 13 days is due to the fact that for some time they did not know that these adjustments are necessary. In general, there will be many corrections over 10,000 years.
Misha promised a grail, but fell into populism 😄
1) Prof. Seveliev says that it is not 100 thousand connections per day that each neuron breaks, but that there are only about 100 thousand of them each, and an average of 3 connections per day are created and broken.
2) A system with re-learning is quite suitable for this. For example, we feed a new batch of retraining data (contradicting previously memorized information) and the coefficient of connection between some neurons recalculated and became zero. This could well be considered an analogue of physical breaking of connection in a real neuron. And another neuron recalculated from 0 to some value >0 - this is an analogue of creating a new connection.
1. Well, I misspoke here, I'm not arguing. I was talking in general about the number of connections between neurons, not one....
2. there is no complete break here. Still zero is transmitted. Although some approximation to this effect is present. Well done!!!!
as a random sequence generator - analog of rand() function
If you put rand() on the input, most likely the "brains" are not enough. I was running lstm on random data for a while and didn't notice anything like that. And if you learn the steps - number systems conversion, algorithm, then it should work.
In general, it is interesting to check on a very powerful iron.
I'm duplicating the video here, just in case anyone would be interested. And I am local in this branch, so what the hell... Maybe some magician there will want to argue, etc.. :-)
https://youtu.be/TlNk3fKkUxo
Ahahahaha )))) GO!!!