"Творческие планы"
Mar. 12th, 2020 06:42 am![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
dmmНесколько вещей, по которым хочется написать нормальные обзоры, и не хватает сил на это. Я их, хотя бы, обозначу, а если будут силы, то тогда возникнут и подробные обзоры.
***
arxiv.org/abs/1905.10985
An essay on AI-generating algorithms by Jeff Clune (22 pages + literature, no formulas)
One of the most important AI-related texts in recent years.
***
The "unusually hardcore meditation book" I mentioned in November: dmm.dreamwidth.org/21712.html
The immediately usable part is Fire Kasina practice, Chapter 29 ( www.mctb.org/mctb2/table-of-contents/part-iii-the-samatha-jhanas/29-kasina-practice/ )
(See also en.wikipedia.org/wiki/Kasina )
It looks like one can obtain a lot of interesting and strong visual effects and cognitive effects by using this technique (safety of that is a whole different topic).
Also there is a wider discussion of Ingram's approach, what one can learn from him, what might be questionable, impact of his book, and, in particular, discussion of relationship between his book and Pelevin's "Тайные виды на гору Фудзи".
***
I finally understood how DeepDream worked. Until now I only understood it very informally ("let's intensify the activity in some parts of a trained neural net, and then DeepDream effects would emerge"):
en.wikipedia.org/wiki/DeepDream
This was not a real understanding (there is currently no known way to define a map from activity of a neural net to what "it might perceive").
What one should do is to define a goal function measuring the activity of the part of neural net one wants to intensify, and to consider the gradient of that goal function with respect to pixels of the input image. And one should change that image according to that gradient, and then depending on what parts of neural nets one wants to intensity and how much, various DeepDream effects appear, as one changes the input image according to that gradient.
Дальше, есть тема для медитации про то, как это соотносится с нашим субъективным визуальным восприятием (assuming that these things are related at all to each other).
***
Вот, я, по крайней мере, обозначил темы, и дал ключевые ссылки.
***
arxiv.org/abs/1905.10985
An essay on AI-generating algorithms by Jeff Clune (22 pages + literature, no formulas)
One of the most important AI-related texts in recent years.
***
The "unusually hardcore meditation book" I mentioned in November: dmm.dreamwidth.org/21712.html
The immediately usable part is Fire Kasina practice, Chapter 29 ( www.mctb.org/mctb2/table-of-contents/part-iii-the-samatha-jhanas/29-kasina-practice/ )
(See also en.wikipedia.org/wiki/Kasina )
It looks like one can obtain a lot of interesting and strong visual effects and cognitive effects by using this technique (safety of that is a whole different topic).
Also there is a wider discussion of Ingram's approach, what one can learn from him, what might be questionable, impact of his book, and, in particular, discussion of relationship between his book and Pelevin's "Тайные виды на гору Фудзи".
***
I finally understood how DeepDream worked. Until now I only understood it very informally ("let's intensify the activity in some parts of a trained neural net, and then DeepDream effects would emerge"):
en.wikipedia.org/wiki/DeepDream
This was not a real understanding (there is currently no known way to define a map from activity of a neural net to what "it might perceive").
What one should do is to define a goal function measuring the activity of the part of neural net one wants to intensify, and to consider the gradient of that goal function with respect to pixels of the input image. And one should change that image according to that gradient, and then depending on what parts of neural nets one wants to intensity and how much, various DeepDream effects appear, as one changes the input image according to that gradient.
Дальше, есть тема для медитации про то, как это соотносится с нашим субъективным визуальным восприятием (assuming that these things are related at all to each other).
***
Вот, я, по крайней мере, обозначил темы, и дал ключевые ссылки.
no subject
Date: 2020-03-12 03:57 pm (UTC)Но исходно это было сделано, чтобы "заглядывать внутрь" искусственных нейронных сетей - понять, где в промежуточных слоях какая кодировка (слои, близкие к входу, учат простые features, так же как в мозгу primary visual cortex areas учат простые features, а в более удалённых от входа слоях более сложные и менее локальные features, и люди изучали, какие именно).
Если начать менять входную картинку по градиенту активности некоторого слоя (берётся производная средней активности, например, одного слоя), то, в некотором смысле, можно сказать, что этот слой как раз кодирует распознавание тех штук, которые начинают возникать на картинке при таком изменении.
no subject
Date: 2020-03-12 04:46 pm (UTC)