Формулировка

[dmm]

Хочется фокусироваться на том, чтобы скульптурно лепить классные штуки из DMMs, и на том, чтобы учить их лепить классные штуки из самих себя.

***

Dataflow matrix machines: more versatile neural machines. It might be useful to be able to easily express algorithms precisely within neural machines, rather than only to learn them approximately, to be able to use readable compact neural networks as well as the overparameterized ones, and to control the degree to which they are overparameterized, to be able to precisely express complicated hierarchical structures and graphs within neural networks, rather than only to model them, and to be able to have flexible self-modification capabilities, where one can take linear combinations and compositions of various self-modification operators and where one is not constrained by the fact that a neural net tends to have more weights than outputs.

It turns out that this can be achieved by a rather mild upgrade: instead of basing neural machines on streams of numbers, one could base them on arbitrary streams supporting the notion of combining several streams with coefficients ("linear combination"). Then one can support the key idea of neural computations, namely that linear and non-linear transformations should be interleaved, and at the same time achieve the wish list above.

***

Здесь возможен  стереовзляд на эту деятельность. Можно подходить к этому всему, как к очень нестандартному варианту программирования, и к использованию машинного обучения для создания систем, умеющих делать этот нестандартный вариант программирования. А можно думать про это, как про нестандартный способ делать разные замечательные виды искусства, особенно динамического искусства, morphing, breathing audio-visuals, and similar things, и к тому, чтобы создавать системы, которые учатся делать такое искусство.

Хочется сохранять этот стереовзгляд, занимаясь всем этим... (Я надеюсь, что всякое довольно стрёмное, что происходит вокруг, тем ни менее, даст ещё некоторое время на усилия в этом направлении; но посмотрим.)

Comments

[dmm]

Надо их ещё как-то чуть больше "антропоморфизировать".

Скажем, когда сеть наращивает сама себя "фрактальным образом", путём клонирования своих собственных подграфов, хорошо бы как-то научиться думать про то, "как она это ощущает изнутри"...

[juan_gandhi]

О. Виды искусства. Создание реальности.

[dmm]

Нарезка из "Synergy between AI-generating algorithms and dataflow matrix machines", covering possible interplay between AI-GAs and DMMs: https://github.com/anhinga/2020-notes/tree/master/research-notes

"Dataflow matrix machines are designed to address [...] problems of "RNNs as programs" and to provide us with a class of neural machines which is expressive enough to constitute a viable programming platform, to host structured information without distorting it by embeddings, to have rich and convenient self-modification facilities, and to be able to encapsulate any algorithms within neurons, as long as those algorithms agree to interface via streams of data for which one can combine several streams with coefficients."

"in the world of DMMs there is no strict boundary between an architecture of the network and the algorithm it implements. Conceptually, we consider a countable address space and a countable-sized network, with only a finite number of weights being non-zero at any given moment of time and the corresponding finite part of the network being allocated in the computer memory and active at that moment of time. The dynamically changing network architecture corresponds to the sparsity structure of the network, i.e. to knowing which weights must be zero and which weights are allowed to be non-zero at the moment; this also corresponds to the notion of program sketch in the world of more traditional programming.

So, in some sense, the total of the first two AI-GAs pillars is structured a bit differently. Instead of distinguishing between learning the architectures and learning the learning algorithms, we distinguish between learning the architecture and learning the weights of a network capable of modifying its own weights and architecture. The primitives inside neurons can be arbitrarily complex (a subnetwork can even be used inside a neuron), so supplying a good library of primitives is an important part of learning the architectures."

[dmm]

"the ability of DMMs to form compact and understandable neural machines with rich functionality is a big advantage, and so is their ability to form modular, hierarchical network structures. It should be much easier to achieve machines which we can understand along this path, compared to the usual mixture of very large traditional neural nets and large amounts of conventional computer code.

The same set of properties of DMMs makes them good construction material for artificial virtual worlds and artificial civilizations. Instead of having people or artificial agents implement artificial worlds in conventional languages, and then define and use complicated APIs, we can simply build those worlds from DMMs. This way, we achieve uniformity of our material in all Three Pillars of AI-GAs, and we have a good degree of control of the extent to which the artificial worlds in question are alive and morphing vs. static and unchanging."

[dmm]

"The prime vector of further DMM development is discovery of new DMM programming idioms and paradigms, and also discovery of new self-modification and machine learning patterns. Currently, we are at the stage when this work is being done manually, but eventually we expect this to become more and more automated.

We do expect a typical AI-GAs advance to be applicable to DMMs in this sense, and to result in novel ways of composing and using DMMs. At some point, we expect the automated processes to match and exceed our own abilities of being creative with DMM design and use. But even before that, a variety of modest AI-GAs-related advances should produce interesting and unexpected DMM-related developments."