Artificial neural networks. - page 8

 
gpwr:

Who established this? We can categorise an object after 50 milliseconds with 80% accuracy. That's 20 objects per second, at any buckgrunge. Many mammals do it even faster to avoid being eaten (evolution). Artificial nets do it in a few seconds, and on an empty buckgrind. The power of the brain is in its parallelism, which we will never be able to achieve by conventional means of computer technology. No one denies the usefulness of automated trading, but networks will not replace the trader's brain in the search for patterns in the market in the next 20-30 years. It takes a lot of neurons. Does anyone here think that a network with 10-20 neurons can replace the trader's brain? What a dumb creature this trader must be!

That's why there are people who work on unusual computer technologies. ))

Kwabena Boahen talks about a computer that works like a brain

Квабена Боахен рассказывает о компьютере, который работает по принципу мозга
Квабена Боахен рассказывает о компьютере, который работает по принципу мозга
  • www.ted.com
Исследователь Квабена Боахен ищет пути повторения в силиконе колоссальной вычислительной мощности человеческого мозга, поскольку изучение беспорядочных, характеризующихся высокой избыточностью процессов, протекающих в голове человека, даёт реальный толчок для создания небольшого, лёгкого и супер...
 
tol64:

So there are people who work on unusual computer technology. ))

Kwabena Boahen talks about a computer that works like a brain

I know Kwabena personally. I also know about the SpiNNaker project from Manchester and its leader Steve Furber, who developed the first ARM(http://apt.cs.man.ac.uk/projects/SpiNNaker/project/). Steve managed to fit 18 ARM processors on a single chip and 48 chips, that is 864 parallel processors. Each processor calculates 500 neurons, i.e. 432 thousand neurons. So far this network does not do anything useful. I am also aware of other groups developing a new type of processor. It is still a long way from reality, hence my prediction that it must wait 20-30 years.

Research Groups: APT - Advanced Processor Technologies (School of Computer Science - The University of Manchester)
Research Groups: APT - Advanced Processor Technologies (School of Computer Science - The University of Manchester)
  • apt.cs.man.ac.uk
What are the Goals of the SpiNNaker Project? SpiNNaker is a novel massively-parallel computer architecture, inspired by the fundamental structure and function of the human brain, which itself is composed of billions of simple computing elements, communicating using unreliable spikes. The project's objectives are two-fold: To provide a...
 
gpwr:

I know Kwabena personally. I also know about the SpiNNaker project from Manchester and its leader Steve Furber, who developed the first ARM(http://apt.cs.man.ac.uk/projects/SpiNNaker/project/). Steve managed to fit 18 ARM processors on one chip and 48 chips, that is 864 parallel processors. Each processor calculates 500 neurons, i.e. 432 thousand neurons. So far this network does not do anything useful. I am also aware of other groups developing a new type of processor. It is still a long way from reality, hence my prediction to wait 20-30 years.

I agree with Joo about "no need to copy nature...".

I also agree with you that the human brain works quite efficiently... but

a person in order to add up " 2 + 2 " needs to recognize the image of "2", then the image of "+" then again "2", then find the association from the "maths" section,

Remember to match the example to the answer.

Don't you think that such a method, although universal (which allowed man to rise above the realm of nature), is ineffective compared to the computer?

In fact, everybody for some reason follows the way of nature, but nature has never had mathematics, and methods tested by nature are not effective in this doctrine, that is why to become an outstanding mathematician one must almost deny the world, and devote himself entirely to mathematics (which in translation means to keep acquired knowledge of mathematics in the nearest associations). But at the same time some dumb machine with MathCad solves everything much more efficiently than the most brilliant mathematician.

SZY imho "computer is a human assistant" like a dog with stronger teeth and a sharper sense of smell.

 
Urain:

I agree with Joo on the "no need to copy nature..." part.

I also agree with you that the human brain works quite efficiently... but

a person in order to add up " 2 + 2 " has to recognise the "2" image, then the "+" image, then again the "2" image, then find the association from the "maths" section,

and then remember to match the example to the answer.

Don't you think that such method, although universal (which allowed man to rise above the realm of nature), but in comparison with the computer is ineffective?

In fact, everybody for some reason follows the way of nature, but nature never had mathematics, and methods tested by nature are not effective in this doctrine, that's why to become an outstanding mathematician one must almost deny the world, and devote oneself entirely to mathematics (which means to keep the received knowledge of mathematics in the nearest associations). But at the same time some dumb machine with MathCad solves everything much more efficiently than the most brilliant mathematician.

SZY imho "computer is a human assistant" like a dog with stronger teeth and a sharper sense of smell.

I'm not following you. The discussion is about artificial neural networks. From my viewpoint, modern artificial networks cannot replace the trader's brain in the search for patterns in the market. So far they only use regression, i.e. modeling of an output (buy/sell) as a non-linear function of inputs. The weights of the network are optimized by minimizing the error in past examples, that does not guarantee its profitability on unlearned data. Increasing number of neurons in the network - like in any other non-linear model - allows to reduce the error in training examples to zero, but it does not help the profitability of the network in the future and only damages it (retraining). Everyone already knows that. In order to ensure that the network has at least some chances, it is necessary to choose such inputs, which have a consistent effect on the output. This choice of inputs is made by us by studying past data and finding regularities. The network itself becomes a tool for non-linear input-output modelling, not for finding patterns. For a network to look for patterns, it must be built according to the principle of our brain. Stupidly increasing the number of neurons in ordinary networks would lead to nothing, otherwise elephants would be as smart as we are (the same number of neurons).

I have not belittled the role of the computer anywhere here, but without a human they would still be iron. It is possible that in the future new types of networks will learn to find patterns in data. But knowing the current state of research in this field, we have to wait and wait. By the way, did anyone ever wonder that science fiction books and films predicted robots in a future that is already in the past but they never came? Mankind has learned to fly to the moon, computers and the internet are fast, but robots are nowhere to be found!

 
gpwr:

I know Kwabena personally. I also know about the SpiNNaker project from Manchester and its leader Steve Furber, who developed the first ARM(http://apt.cs.man.ac.uk/projects/SpiNNaker/project/). Steve managed to fit 18 ARM processors on one chip and 48 chips, that is 864 parallel processors. Each processor calculates 500 neurons, i.e. 432 thousand neurons. So far this network does not do anything useful. I am also aware of other groups developing a new type of processor. So far it is far away from reality, which is why I predicted to wait 20-30 years.

It is great that you know such researchers personally. Do you happen to know Henry Markram? His prediction in 2009 was 10 years. :) I wonder where he stands now.

Henry Markram is building a brain in a supercomputer

Генри Маркрам строит мозг в суперкомпьютере
Генри Маркрам строит мозг в суперкомпьютере
  • www.ted.com
Тайны устройства разума могут быть решены, и довольно скоро, говорит Генри Маркрам. Поскольку умственные заболевания, память и восприятие составлены из нейронов и электрических сигналов, он планирует обнаружить всё это с помощью суперкомпьютера, который смоделирует все 100...
 
gpwr:

I do not understand you. The discussion is about artificial neural networks. My point of view is that modern artificial networks do not allow to replace the trader's brain in the search for patterns in the market. So far they only use regression, i.e. modeling of output (buy/sell) as a non-linear function of inputs. The weights of the network are optimized by minimizing the error in past examples, that does not guarantee its profitability on unlearned data. Increasing number of neurons in the network - like in any other non-linear model - allows to reduce the error in training examples to zero, but it does not help the profitability of the network in the future and only damages it (retraining). Everyone already knows that. In order to ensure that the network has at least some chances, it is necessary to choose such inputs, which have a consistent effect on the output. This choice of inputs is made by us by studying past data and finding regularities. The network itself becomes a tool for non-linear input-output modelling, not for finding patterns. For a network to look for patterns, it must be built according to the principle of our brain. Stupidly increasing the number of neurons in ordinary networks would lead to nothing, otherwise elephants would be as smart as we are (the same number of neurons).

I have not belittled the role of the computer anywhere here, but without a human they would still be iron. It is possible that in the future new types of networks will learn to find patterns in data. But knowing the current state of research in this field, we have to wait and wait. By the way, did anyone ever wonder that science fiction books and films predicted robots in a future that is already in the past but they never came? Flying to the moon, fast computers and the internet but no robots!

I simply questioned the very direction of research in NS, the very paradigm of copying nature.

I have great doubts that a network built in the image and likeness of the human brain will surpass the creator.

I believe that NS research needs to move in the direction of direct perception of digital data, whereas now the numbers for NS are just images.

 
Urain: I simply questioned the very direction of research in NS, the very paradigm of copying nature.

technical means have never copied nature, be it a wheel or an aeroplane, but are perfectly capable of doing their job, so NS should work with mathematical models and should not imitate analytics/decision making of a trader

SZS: Imagine what a camera would look like that replicates the process of an artist )))))

 
IgorM:

technical means have never copied nature, be it a wheel or an aeroplane, but are perfectly capable of doing their job, so NS should work with mathematical models and should not imitate analytics/decision making of a trader

ZS: Imagine what a camera would look like that replicates the process of an artist )))))

The camera copies the eye, so the example doesn't count. But in general you got the point of my post right.
 
Urain:
The camera copies the eye, so the example doesn't count. But overall you got the point of my post right.
But I also made an argument against it. The camera many times surpasses the capabilities of the eye, if it is also a telescope. ))
 
Urain: The camera copies the eye, so the example doesn't count. But in general you got the point of my post right.

I agree that the camera copies the eye, but the result is the same as for an artist - an image on paper, the only thing I did not describe the technological process of making a photo

So we finally understood why the NS are not always successful in trading: the problem is not with the NS, but with the mathematical model of market information, which the NS provide for training - who close the last 2, 3, ... 100 bars, who read the technical indicators, in a word "for what you're good at", we have to understand what market information is really important for trading - patterns? Last few bars? Volumes? Time of day? .... And the sad thing is that having filtered out unnecessary information and created a mathematical model of the market, you can build an effective TS without NS.